AI Studio User Manual

1. Introduction & quick start

AI Studio is an AI video and image generation workspace for creators — generation, editing, and asset management all in one place.

Here is what you can do with it:

• AI video generation: turn text, a single image, or multiple reference assets into short videos (Seedance, Kling, MiniMax H3, HappyHorse, Veo, Gemini and more).
• AI image generation: generate images from text, or repaint uploaded images (GPT Image 2 / 2.5, Nano Banana Pro / 2, Seedream 5 Pro).
• Smart edit: annotate a generated image by hand (markers, arrows, text) and let AI apply the changes in place.
• Storyboard: upload product / character / scene material and get a multi-shot storyboard, then render it into a marketing video.
• Product Listing Set: upload product photos and let AI plan and batch-render a full set of e-commerce listing images (white background, lifestyle, feature callouts), with 20 built-in design style templates; also plans Amazon A+ Content sets against the official module specs.
• Image Replica: supply style reference images plus your own product photos, and AI batch-produces listing images that follow the reference layout.
• Garment Pattern Variations: upload 1–6 reference images (garment shots, pattern swatches, worn photos — any mix, no grouping needed); AI freely auto-identifies each image's role plus an overall style and selling positioning, and one submit after you confirm the details drafts the plan and renders every flat-lay automatically — approve them, then batch-generate the model shots with one click.
• Pattern Transfer: upload the same style in several colours (1–12 photos) plus one pattern (direct upload, or free extraction from a viral photo); AI migrates the pattern pixel-faithfully onto every colour and renders a flat-lay plus a model shot for each — one model throughout, a different scene for every colour.
• Text to Speech: type a script, pick a curated voice and a model tier, and synthesize natural spoken audio — with pause markers, emotion tags (Expressive tier) and AI script polish.
• Motion Transfer: upload one character photo plus one motion video and the subject in the photo (a person or a pet) performs the video's motion — optionally in the video's own setting.
• Video Edit: modify / add / remove elements, edit sound, extend and splice an existing video — upload it or bring it in from any result card, then describe the change in one sentence.
• Batch Video Remake (Batch Production): one viral reference video + character / scene photos + the product images of a batch of same-style SKUs — one click renders a matching video per SKU with no interaction after launch.
• Voice Clone: upload 1–3 clean reference recordings and clone your own voice, then pick it right inside Text to Speech (visible to you only).
• Viral Video Remake: upload a viral reference video; AI analyses its camera work and voice-over rhythm, swaps in your own character / product / scene, re-renders segment by segment, and stitches the final cut.
• Video Upscale: turn low-resolution footage into a sharp 720p–4K version.
• Work & asset management: every result lands in My Creations, and everything you upload is kept in the asset library for reuse.

Quick start in three steps

1. Open the site, click Sign in in the top-right corner and register or sign in with an email address or Google.
2. Enter the Studio dashboard and pick a tool from the left menu (Text to Video, Text to Image, Storyboard, …).
3. Add reference materials by clicking Add, reusing from the Library, or simply dragging files from your desktop onto the material area (the zone highlights; mismatched types are skipped with a notice).
4. Write your prompt, choose a model and parameters, then click Generate. Credits are charged when you submit; the result appears in the preview panel and can be downloaded.

2. Sign up & sign in

The sign-up page offers Sign up with code (default, passwordless) and Sign up with password; the sign-in page offers Sign in with password (default) and Sign in with code, with Google sign-in kept available.

2.1 Email sign-up and sign-in

Sign up with a code (recommended, no password needed):

1. Stay on the Sign up with code tab, enter your email (name optional) and click Send code.
2. Open your inbox for the 6-digit code (valid for 5 minutes; resend after 60 seconds), enter it and click Create account & sign in.
3. Once the code checks out you are registered, verified and taken straight to the studio. Next time use the Sign in with code tab; if you ever want a password, set one via "Forgot password" on the sign-in page.

Sign up with a password:

1. Switch to the Sign up with password tab, fill in name, email, password and confirm password (both entries must match), then submit.
2. A verification email is sent — open it and click the confirmation link; the original sign-up page then continues to the studio automatically (clicking the link on another device signs you in there as well).

To sign in:

1. Open the sign-in page; the email/password form appears directly.
2. Enter your email and password; the eye icon toggles password visibility.
3. Click Sign in. Forgot your password? Use Forgot password? above the password field.

2.2 Sign in with Google

Click Sign in with Google and authorise the account in the Google consent screen.

3. Dashboard overview & navigation

Every sign-in (email, phone code or Google) lands on the Studio dashboard: the tool menu on the left, your own data dashboard in the middle (balance, output, running tasks, spend and failures), recent work at the bottom. There are no tool cards any more — every tool is reached from the left menu.

The data dashboard (since 2026-09-12)

The home page no longer lists tool cards — every tool lives in the left menu, and the home page shows your own numbers. Dark is the baseline theme and light is fully supported; the layout works on desktop and phones.

Left menu

Top and bottom areas

• Account area (bottom left): your name and credit balance. Click it to switch accounts or sign out.
• Top right: language (EN / 中文) and theme (light / dark) switches.
• Recent work: the dashboard home lists recent results; View all opens My Creations.

4. How credits work

Generating images and videos consumes credits. Knowing the rules helps you plan your work.

Where credits come from

• Where credits come from: new accounts receive a 50-credit sign-up bonus (since 2026-09-12; only the first account per sign-up IP gets it, shown as "Signup bonus" in billing), and further credits come from topping up.
• Top-up purchase (launched 2026-08): open Pricing from the top navbar or the sidebar. Four one-time tiers are available: ¥250 for 5,000 credits, ¥500 for 10,500 (+500 bonus), ¥2,500 for 55,000 (+5,000 bonus), ¥5,000 for 115,000 (+15,000 bonus). Payment is currently by corporate bank transfer: click "Top up now" and contact 路易电商 (Louis E-commerce) as prompted, including your account email in the transfer note. Top-up credits never expire, and top-ups appear in the billing history.

How credits are spent

1. Every task shows its exact price on the submit button before you commit (for example “Generate · 200 credits”).
2. Credits are charged up front. If your known balance is too low, the paid action stays clickable and shows a dedicated “Insufficient generation credits” message without starting a task. The server checks again when you submit, so the same message also covers a balance changed in another tab.
3. The price depends on the model, resolution / quality, duration and number of reference images — higher settings cost more. The number on the button is always authoritative.

Team accounts (since 2026-09)

A company or team can top up one owner account and let several member accounts use those credits. Entry point: sidebar Settings › Team. There are two credit modes, chosen by the owner in the "Credit mode" card at the top of the Team page: Allocation (the owner allocates on demand and can reclaim) and Shared pool (members spend the owner's credits directly, since 2026-09-04).

1. Invite a member: on the Team page the owner enters the member's email and sends an invitation. The member receives an email; the page also shows a one-time invitation link you can copy and share directly (valid for 7 days — afterwards use "New link" or revoke it under pending invitations).
2. Accept: the member opens the link, signs in or signs up with the invited email address (email-code sign-up works without a password) and clicks "Join team". A member belongs to exactly one owner, and an owner cannot join another team as a member.
3. Allocate and reclaim: in the members table the owner allocates credits to a member (they land in the member's own balance) or reclaims what is unused. The reclaimable amount is the smaller of the member's current balance and the owner's net allocation, so sign-up bonuses and other credits that were never the owner's are never taken back.
4. Usage: a "Member usage" bar chart at the top of the members card ranks every member's consumption for the selected window; the owner can also review each member's credit usage over the last 24 hours / 7 days / 30 days — counted only from the moment the member joined the team, never earlier spending — the breakdown by module and the individual ledger entries (time, type, module, model, credits, balance). The owner never sees a member's creations: prompts, uploads and results stay private to the member.
5. Remove a member: removal automatically reclaims the member's reclaimable credits; the account becomes a regular account again and the ledger history is kept.
6. Shared-pool mode: once switched, members no longer need allocations — their spending is charged to the owner's balance; a member's balance reads "Team credits · master X" and shows the team's remaining credits; the owner sees each member's consumption on the Team page, and ledger rows for members' spending name who spent. Switching to the shared pool leaves credits already allocated frozen with the members; switching back to allocation makes members use their own balance immediately. When the pool runs out, members see "The team's credit pool is insufficient. Ask the master to top up."

Processing time and timeouts

5. AI video generation

Three modes share one interface: Text to Video, Image to Video and Reference to Video. Switch between them from the left menu.

5.1 Text to Video

Generate a short video from a written description — ideal for drafting ideas quickly.

1. Open Text to Video from the left menu.
2. Model: pick a video model from the dropdown (see 5.4). Each entry carries its brand icon.
3. Resolution: 480p / 720p and so on, depending on the model.
4. Duration: drag the slider or click a duration button.
5. Aspect ratio: 16:9, 9:16, 1:1 and others.
6. Advanced: expand for the audio toggle and, on some models, a random seed.
7. Prompt: describe the shot, camera movement, lighting and mood.
8. Check the price on Generate · N credits and submit.

5.2 Image to Video

Upload one image as the first frame and let AI bring it to life.

1. Open Image to Video.
2. Click Add in the first frame slot to upload, or Library to reuse something you uploaded before.
3. Some models also accept a last frame; single-frame models only show the first slot.
4. Set model, resolution, duration and aspect ratio.
5. Describe the motion you want (“slow push-in”, “the subject turns their head slightly”).
6. Click Generate.

5.3 Reference to Video

Supply several reference images — plus reference video and audio — to keep a character, product or style consistent across a series.

1. Open Reference to Video.
2. Upload material into the reference image / video / audio groups. Each group header shows its quota (for example “0/9”), which varies by model. Hovering the reference-image upload area floats the Reference image tips guide (good images get green checks / bad ones red crosses: clean backgrounds, real model shots, close-up details, one-angle-per-image, showing the real size and a brand-text close-up work best; collages/grids, promo-text or watermark overlays, picture-in-picture shots, tiny cluttered products, mixed variants and low-res blurry shots make the model guess and mismatch your product; for people use one frontal portrait plus an optional full-body shot — never a multi-view sheet, and keep other people's clear faces out of worn shots) — it shows on every hover; on touch devices use the Image tips entry beside the group title.
3. Type @ in the prompt to open the list of uploaded material; picking one inserts a mention chip that refers to that exact asset.
4. Set the model and parameters. For Seedance the generation route is scheduled automatically by the platform — nothing to choose. Then click Generate.

5.3+ Prompt wizard (Seedance family)

Not sure how to write a video prompt? In Reference to Video with a Seedance-family model or MiniMax H3 selected, a Prompt wizard button appears beside the prompt box (for H3 the draft follows its official English six-section format) (not offered in Image/Text to Video for now). The wizard supports category switching (Apparel and General at launch, Apparel default; media analysis detects the category automatically and never overrides a manual pick) — detail-mining vocabulary and suggested actions adapt per category, with more categories to come. It is a single-page form: an editable analysis card plus six sections (asset roles → goal → scene & mood → subject & action → camera & rhythm → audio) on one screen, always visible and editable. A progress rail on top lights up completed sections, highlights the suggested next one, and jumps on click; fill them in and generate a structured prompt following the selected model's official prompt guide. Closing and reopening keeps your input; Reset clears it.

1. Every step offers suggested options plus free-form input. When images are uploaded, the wizard first runs a free AI media analysis (product, selling points, audience); suggestions come from it, and the analysed product/selling points/audience prefill the editable card on top — one edit makes it the authoritative fact, and edits ride into the drafting call.
2. The asset-roles section assigns a purpose to each image/video: single-select for images (product reference, subject reference, …; the default “Auto-assign” lets the AI pick roles from what each image actually shows — the reference images are attached to the drafting call). Videos are multi-select and come pre-checked as “Motion reference + Camera reference” (leave it and the source is mirrored shot by shot); a video marked style-reference only is not shot-mirrored — only its mood and pacing are borrowed. You can also explicitly mark an asset as “Not used”. Every role actually takes effect in the final prompt: first frame = shot 1 starts from that image, last frame = the final shot settles on it, props are reproduced as pictured and actually used in the action, identity photos contribute face/hair/build only (wardrobe from product refs), scene images strictly define the environment.
3. With a reference video: one-click analysis. Once the roles are confirmed, a “Analyze the reference video to draft the plan” card sits between the assets and goal sections (it waits while the media analysis is still running, then becomes clickable). One click breaks the video down by the confirmed roles and fills in at once the goal tag, the scene & mood (the source environment is restored when no scene image is given; with a “Scene & style reference” image it describes that image instead), the per-shot action plan and the camera plan — every section stays editable, and you can regenerate. With several selected shots (up to 3), the wizard joins them in order into one reference timeline and analyses segment by segment with source labels such as “4-8s (video 2): …”. The breakdown only restores the source; drafting then swaps the source’s people and objects for your subject and product — the source cast’s wardrobe is never copied, and every shot restates the product’s colours and structure. A static display shot of the source product is not mirrored — only its framing is borrowed and the displayed items are rewritten from your product images, without forcing a product image as the first frame (the shot picker flags such shots and suggests leaving them unchecked). With a reference video, assign at least one image as Product reference; the wizard reminds you when none is set. For brand lettering, upload a lettering image and assign it the Logo-reference role (the wizard prompts you automatically; lightweight models are unreliable at letters — use 2.5 or 2.0 Standard). Below the Audio section, a free-form Extra-requirements box takes anything else and the draft fulfils every item. Product-swap mode: when the source is detected as a same-category product demo, the shot picker opens in swap mode — pick one contiguous span (4-30s, Seedance 2.5) and the result keeps the source people, motion, camera and cuts while swapping the objects for your product; the wizard maps source objects one by one (replace / remove / keep) instead of reskinning them. The breakdown also writes an identity card from the subject-reference image (including how the hair looks from behind), and after linting the draft gets a mechanical "Identity lock · top priority" block: one subject throughout, every angle follows the subject image, back shots keep that hairstyle as seen from behind, people in the reference video and product images are only motion sources or mannequins, restated on every turning shot; the block is visible and editable in the draft (not added in product-swap mode or with H3).
4. Without a reference video: serial cascade. Pick a goal, generate scene & tone ideas (matched to the goal and product), confirm the scene, then the AI drafts three complete timestamped action plans designed inside that scene (choose one, editable), and finally derives three camera schemes from the chosen action — each step builds on the previous confirmation. Assigning an image the “Scene & style reference” role makes the shooting environment strictly follow that image. The Scene / Action / Camera and Audio sections each carry a notes-then-generate bar (“Section notes” input with the generate button beside it; Enter also generates): state your constraints before generating (e.g. “must include an unboxing moment”) and the plans are designed around them from the start; edit the notes and hit Regenerate (or Enter) to redraft.
5. The audio section decides the clip's sound design: pick Voice-over to generate three per-shot, time-segmented VO scripts (lines in curly braces, pacing matched to each shot's length, product facts drawn only from your confirmed product info; the VO language is selectable — 中文, English, Español, 日本語, ไทย, Bahasa Melayu — defaulting to the site language); or Music only for three per-segment music plans; with a reference video you can also pick Follow source audio. Choose one of three and keep editing; the final prompt embeds the lines per shot in the official curly-brace syntax, or writes the music into each shot's audio slot. Leave it unselected to add no audio constraint. Note: the sound design only takes effect when the Generate audio toggle is on.
6. Click Generate prompt (10 credits per generation, fully auto-refunded when you apply it and submit a video render within 24 hours — effectively free when used here; the questions, media analysis and video analysis are free). The result is editable, and Apply fills it into the prompt box — asset references become mention chips, and everything downstream works exactly like a hand-written prompt.
7. Prompt health check and Fix all. A "Prompt health check" panel appears under the draft: every asset referenced, no missing or "not used" assets cited, product shots spelling out colour/material, no "toddler/blogger"-style aliases replacing the defined subject, time segments covering the full duration, voice-over length fitting the duration, and no render parameters such as 4K/16:9 in the body. All-green means it passed; otherwise each finding is listed — edit the text directly (re-checked as you type) or click "Fix all · 10 credits" to patch only the named items while everything else (including your edits) stays verbatim; no charge if nothing could be improved (rebated once applied and rendered). The server runs the same check plus one or two patch rounds during generation.
8. Video feedback. Under every successful result: "Did this video meet your expectations? 👍 👎" — 👍 submits at once, 👎 lets you tick what went wrong (product looks wrong / person changes / wrong action, camera or scene / voice-over or audio / subtitles or watermark / quality or distortion / pacing) and add a note, editable any time. It never affects the task or credits; feedback is linked to the wizard recipe you used and drives the next round of wizard and prompt improvements.

5.4 Video model reference

5.5 Viewing & downloading results

1. After submitting, the preview panel shows “Waiting / Generating” and then the finished video.
2. Click to play; use Download to save it locally, or Save to library to copy the finished video into your asset library (free; pick a folder when you have folders, and a repeat click reports it is already there). The copy can then be picked from the library in the video slots of Reference to Video, Video Remake, Motion Transfer and Batch Video Remake. Deleting the original task does not affect the saved copy.
3. Every task also appears in My Creations (chapter 17) where you can reopen or retry it.

6. AI image generation

Two modes share one interface: Text to Image (prompt only) and Image to Image (repaint an uploaded reference).

6.1 Text to Image

1. Open Text to Image from the left menu.
2. Model: pick an image model (see 6.3).
3. Aspect ratio: Auto / 1:1 / 3:2 / 16:9 / 9:16 and more; some models add a Custom size.
4. Quality / resolution: GPT Image 2 Flex offers auto / low / medium / high quality tiers; other models offer 1K / 2K / 4K.
5. Background: the GPT Image 2.5 lines add a three-way Background switch (Auto / Opaque / Transparent); Transparent outputs a PNG with an alpha channel.
6. Prompt: describe the image.
7. Check Generate · N credits and submit.

6.2 Image to Image & the asset library

Upload one or more reference images and let AI repaint them.

1. Open Image to Image.
2. Click Add to upload, or Library to reuse an earlier asset.
3. Describe the change you want (“replace the background with a clean white studio”).
4. Set the model and parameters, then click Generate.

6.3 Image model reference

6.4 AI prompt writer

Not sure how to phrase a prompt? Both Text to Image and Image to Image have an AI prompt writer · 2 credits button at the top right of the prompt box. Write one line describing the result you want; the AI combines your reference images with the official prompting style of the selected model (GPT Image 2 / 2.5, Nano Banana and Seedream each phrase prompts differently) and writes a professional English prompt for you.

1. Pick the model, aspect ratio and (for Image to Image) the reference images first, then click AI prompt writer above the prompt box.
2. The dialog lists the current reference images; describe the result in one line under “What you want” (e.g. “white-background hero shot, product centered, soft studio light”). Any text already in the prompt box is carried over.
3. Click Write prompt · 2 credits. After a few seconds you get an English prompt plus a one-line summary of the key points; the prompt is editable right in the dialog.
4. Click Rewrite for another attempt (another 2 credits), or Apply to prompt to fill the prompt box and generate as usual.

7. Smart edit

Make targeted changes to a finished image: draw markers, arrows and notes on it, describe what should change, and AI produces a new image from your annotations. An edit never overwrites the original — the pre-edit image stays in the version history and you can switch back at any time. Works on Text to Image and Image to Image results, storyboard panels, and every card in Product Listing Set, Image Replica, Garment Pattern Variations and Pattern Transfer.

Where to find it

Once an image finishes, a Smart edit button appears next to Download in the preview panel. Storyboard boards have the same button in their action row, and listing / replica / garment-pattern / pattern-transfer cards carry their own entry point.

Version history

After a successful edit, a Version history strip appears under the preview: thumbnails of the Original and every Edit N, with the version currently in use marked Current. Click a thumbnail to zoom; every version can be downloaded or saved to the library, and Set as current on any other version swaps it back into the result slot (free, nothing is regenerated) so later downloads, saves and further edits start from it. On Product Listing Set / Image Replica / Garment Pattern Variations / Pattern Transfer cards the same list opens from a History N button; a storyboard board's paid re-render also enters the history, and restoring such a version brings back the script it was rendered from. Up to 20 versions are kept per image — the oldest edits are dropped beyond that, the original never is.

Color lock (Image to Image / Smart edit)

A model repaints the whole frame with its own colour bias (typically a red or magenta cast), and feeding a result back in as the next reference stacks that cast round after round. Image-to-image and smart-edit results now carry a Color lock action: it corrects the overall cast against the reference (the first reference image's original for image-to-image, the pre-edit version for smart edit) and then restores every untouched area from the reference's own pixels, keeping only what actually changed. No model call, no credits, a few seconds; the result enters the version history as Color-locked N, set as current and free to switch back from.

• Auto (on by default) runs it once when a result lands or a smart edit finishes; opening an older task never triggers it — click the button when you need it.
• Only aligned results qualify: scene or pose changes that repaint the composition are skipped with a notice, as are results whose colours already match the reference.
• When the reference is much smaller than the result (under 90% of its size) or the result has a transparent background, only the cast is corrected — pixels are not restored.
• The notice reports the colour difference before and after plus the share of restored pixels; a version that is already colour-locked is not locked again, but you can switch back to the original and run it once more.
• Needs desktop Chrome / Edge / Firefox or Safari 16.4+; unsupported mobile browsers show a notice.

Editor tools

Steps

1. Click Smart edit to open the editor.
2. Annotate the areas you want changed.
3. Describe the change in the box at the bottom; you can reference marker numbers (“replace the red apple at marker 1 with a green one”).
4. Pick the model and resolution for this edit (the original image's model is preselected) and check the price.
5. Click Edit · N credits. The image shows “Editing…” and the new version replaces the original when it finishes.

8. Storyboard (AI Agent)

A complete short-video pipeline for e-commerce: upload product / character / scene material → an AI scriptwriter writes the board prompt → an image model renders a multi-shot board → the AI writes the video prompt → a video model renders the final cut.

Steps

1. Upload material: add references to the product (required), character and scene groups. The three groups share one quota (8 images on most models, 6 on Gemini Omni), so you can split it however you like.
2. Video parameters: target aspect ratio, voice-over language (English / 中文 / Español / 日本語 / Bahasa Melayu / ไทย), prompt model, board model, video model, resolution, duration and shot count.
3. Extra requirements (optional): selling points, style, target audience — the AI gives these priority.
4. Generate: the scriptwriter writes the board prompt, then the board model renders the panels. The prompt box on the right fills in live and stays editable.
5. Video prompt: once the board is ready, click AI Assist in the video-prompt block to have the AI write the video prompt from the board.
6. Render: confirm and submit — the video model renders the final cut.

Parameters

9. Product Listing Set (AI Agent)

Upload product photos and let AI plan and batch-render a complete set of listing images: white-background hero shots, lifestyle scenes, feature callouts and more. Built for Amazon, TikTok Shop, AliExpress and similar marketplaces.

Steps

1. Product photos: up to 8, uploaded or picked from the library.
2. Platform / market / copy language: choose the marketplace (TikTok Shop, Amazon listing set, Amazon A+ Content, AliExpress, own store, …), the target market and the language of on-image copy. The AI plans against each platform's real image conventions (Amazon's pure-white background and ~85% frame fill, for instance); choosing Amazon A+ Content switches to A+ mode (see the end of this section).
3. Selling points: fill in the five-line skeleton (product name / key selling points / target audience / intended scenes / size & specs), or click AI generate (2 credits) to have the AI draft it from your photos and then edit it.
4. Design style — pick one of three routes:
   • AI pick: click AI style analysis and the AI proposes a style from your photos (2 credits); the text stays editable.
   • Style templates: choose one of 20 curated presets (Figure 9-2) across four groups — premium, natural, vibrant and category-specific — such as “Premium editorial”, “Creamy pastel”, “Dark luxury” or “UGC casual”.
   • Custom brief: type your own requirements: palette, layout, creative direction.
5. Set structure: Smart match lets the AI decide how many images of each type; Custom lets you set the counts yourself (4–12 images in total, at least one white-background shot).
6. Generate the plan (2 credits): the AI plans a category, title and English render prompt for every image and lists them as cards. Cards can be selected, deselected and edited.
7. Batch render: choose the image model (GPT Image 2 / 2.5, Seedream 5 Pro, Nano Banana Pro / 2), quality tier and aspect ratio on the right, then click Generate selected · N credits. Each image is charged and rendered independently and appears as soon as it is ready.

Per-image actions

Amazon A+ Content set

Pick Amazon A+ Content as the target platform to switch into A+ mode: the set is planned against Amazon's official standard-module specs across six categories, and every card ships with suggested module copy you can paste straight into Seller Central.

• Six categories map to official modules: Brand hero (StandardHeaderImageText, ≥970×600), Wide banner (StandardImageTextOverlay, ≥970×300), Feature blocks (Three Images & Text, ≥300×300), Lifestyle scene (StandardImageSidebar, ≥300×400), Comparison columns (StandardComparisonTable, ≥150×300 per column) and Specs detail (SpecsDetail, ≥300×300). The default mix is 7 images in exactly 5 modules (basic A+ allows at most 5 modules per ASIN), with at least one brand hero.
• Minimal-text images + module copy: images stay text-free by default — selling points are proven visually (Amazon reviews text embedded in images against its banned-keyword lists, and small text that is illegible on mobile gets content rejected) — while every card also outputs a suggested module headline (≤160 chars), body text (≤1000 chars) and alt text (image keywords, ≤100 chars — required at submission, read by screen readers and indexed for search) with one-click Copy headline / Copy body text / Copy alt text buttons for Seller Central.
• Model and ratio are yours to pick: exactly like the listing set, you choose the render model, quality tier / resolution and aspect ratio (custom size included) at render time; each category row in the structure config shows the suggested ratio closest to its official module (hero 16:9, banner 3:1, feature/specs 1:1, scene 3:4, comparison 1:2) for guidance, and the module-size download centre-crops to the official pixels — the plan's compositions already leave crop margins. Keep renders at or above the module minimum (a 1K 16:9 render is slightly short of 970×600; 2K is safer).
• Module-size downloads: finished cards offer a download cropped to the exact official module pixels (970×600 and so on); the Amazon listing set's white-background main image also gains a 2000×2000 direct download.
• Compliance built in (aligned line by line with the official A+ Content Guidelines): planning automatically avoids the official prohibitions — pricing/promo and call-to-action wording, warranty claims, satisfaction and "best seller"-style boasts, unsubstantiated certification/award/eco/medical claims, contact details and hyperlinks, competitor references, time-sensitive wording and seller-identity phrasing never enter the images or the suggested copy; comparison shots only compare your own variants, before/after use, or the officially allowed unbranded category comparison (LED vs halogen bulbs).
• AI-person disclosure: if a finished image contains photorealistic AI-generated people, tick "AI-generated persons" when uploading it in A+ Content Manager (or tag the file with contains-synthetic-performer metadata) — Amazon requires this disclosure, and the delivery area shows a matching note.
• Review expectations: Amazon reviews A+ submissions in up to 7 business days and publishes approved content within 24 hours; if rejected, follow the official best practices — check image sharpness and remove claim/pricing/contact wording.

10. Image Replica (AI Agent)

Found a listing image whose style you love? Upload it as a style reference, add your own product photos, and AI batch-produces listing images in that style — one output per reference image.

Steps

1. Style references: up to 12. Every reference produces one output in its style.
2. Product photos: your own product. The cap adapts to the model and the number of character photos and is shown live in the UI. Only your product is ever painted — never the one in the reference image.
3. Character photos (optional): up to 2. Only the face, hair and build are taken from them — the garment always comes from your product photos, which is what makes this work for apparel.
4. Market / copy language: drives on-image copy and the visual sensibility.
5. Fidelity — pick one: Layout reference keeps the layout, background structure and subject relationships but designs the palette around your product; Strict replica reproduces composition, layout, palette and detailing, swapping only the product and the selling points.
6. Key selling points (required): type them or click AI generate (2 credits).
7. Unified brief (optional): constraints for the whole set, e.g. “all copy in English”, “keep the model's pose unchanged”. Each requirement is written into every card's prompt: numbers (sizes / thickness / angles) verbatim, style names and structural selling words in explicit English; numbers are also machine-checked and auto-repaired once if missing.
8. Generate the plan (2 credits): the AI writes one render prompt per reference; cards are selectable and editable.
9. Batch render: choose model, quality tier and ratio on the right, then click Generate selected. Per-card actions (smart edit / save to library / download / rewrite & regenerate) match Product Listing Set.

11. Garment Pattern Variations (AI Agent)

Upload 1–6 garment, pattern, worn or detail references into a local draft. Click Start analysis (free) to create a durable task, review the saved AI interpretation, then click Confirm generation to pay for the plan and flat-lays. Once those finish, batch-generate the model shots with one click.

Steps

1. Build the local draft (1–6 images, any mix, no grouping needed): garment shot (the source of truth for cut and silhouette), pattern swatch (style DNA), worn photo (cut plus pattern), or detail reference (context only). Add, remove and reorder freely. Uploading alone makes no AI call and creates no task.
2. Three optional strict reference groups (new 2026-08-31): below the source area you can also upload a flat-lay base (1 image — flat-lay renders reproduce its exact surface, backdrop and props, replacing the default clean/lifestyle styling), model references (up to 2 — model shots use exactly this person's face, hair and build, while the worn garment always comes from the flat-lay; best with one frontal portrait plus an optional head close-up), and a scene reference (1 image — model shots are staged in exactly this environment). All three are optional: leave any empty and AI matches it to the product; they stay editable during the review step. Extra references count toward that render's input-image total (Seedream adds its small per-input surcharge; GPT / Nano Banana flat pricing is unaffected).
3. Start analysis (free): this deliberate click creates the task first, then analyses every source in the durable background workflow. The task immediately appears in My Creations; you can leave, close the browser or switch devices without interrupting it. No credits or billing entry are created.
4. Review the saved analysis (single product, since 2026-08-31): one task fissions ONE product — the full product card (role + summary + analysis fields + variant count) attaches to the single pattern source image only (the first pattern-bearing image by default); every other image collapses to a compact row with a read-only role badge, and its Use as pattern source button moves the whole product card (and the variant count) onto that image instead. A degraded source is marked Needs manual confirmation. Source images and order stay frozen, while the product card's role, summary and analysis fields remain editable.
5. Category / target wearer / base colour: AI fills these in from the uploads; all editable. Base colour is optional free text (e.g. "black", "off-white").
6. Style DNA / selling positioning / suggested scene / suggested model persona: four AI-drafted notes (colour & motif, target audience, shoot setting, model styling) — edit any of them directly.
7. Variant count (single product): the count belongs to the single pattern-source image only (default 10, stepper 0–20); the other images serve as the version anchor or detail references and no longer carry counts of their own.
8. Model / tier / both ratios: Garment Pattern Variations supports GPT Image 2.5 Sunburst (default), GPT Image 2.5 Flare, GPT Image 2 HD, Nano Banana Pro, Nano Banana 2, and Seedream 5 Pro; choose the quality tier plus a separate flat-lay ratio and model shot ratio (both default to 1:1). These settings freeze only when paid generation is confirmed.
9. Confirm generation (2 credits for the plan, plus every variant's flat-lay priced by the chosen model — the exact total is shown on the button): this is the first paid action. AI drafts the plan and renders every flat-lay immediately.
10. English Logo and typography: each new variant receives a distinct fictional English brand name, optional slogan and matching typography/layout. Prompt regeneration preserves its frozen design; model shots copy the latest flat-lay's exact visible spelling and layout.
11. Results (three-column workflow): each variant is one row, showing Flat-lay / Model shot / Pattern vector file in generation order — three columns side by side on desktop, the same order stacked vertically on phones. Each card header has a checkbox — checkable once that card's own flat-lay succeeded, checked by default.
12. Multi-select batch generation: once every flat-lay has finished (success or fail) and at least one succeeded, a toolbar appears at the top of the results — a select-all checkbox (toggles between select-all and clear-all) plus two independent buttons, Generate selected model shots and Generate selected pattern vectors, each priced live from the cards that are both checked and don't already have that render. Uncheck a card to leave it out of this batch; a card that already has a model shot or pattern vector is never re-charged or re-rendered by either button, even while still checked — the two buttons' eligibility is independent. You can also click Generate model shot on a single card.
13. Pattern vector file (print-ready, the third result column): once a variant's flat-lay renders successfully, that column shows a Generate pattern vector button (8 credits, regardless of whether a model shot exists) — AI extracts only the printed graphic from that flat-lay and redraws it as a standalone file: white background made transparent, roughly 4096px wide (about 34.7cm at 300dpi, inside the 32-36cm print-size range). The preview renders over a checkerboard backdrop so the transparency is visible; Download saves a PNG with a -300dpi filename suffix. A failed render can be retried in place at the quoted price.
14. Unified edit and regeneration: the separate regeneration buttons below each image are removed. Use Edit & regenerate at the top-right of the prompt area to edit the flat-lay prompt, model persona, shooting scene and lighting. Changing only model settings regenerates only the model shot. Changing the flat-lay prompt first regenerates the flat-lay, then automatically renders a new model shot from it. The submit button shows the actual scope and credits; the durable two-stage job continues if you leave the page.
15. Other per-card actions: Smart edit (chapter 7 — annotate and overwrite in place), zoom, download and Save to library (a free independent copy that survives task deletion; with folders you pick which one to file it under).

12. Pattern Transfer (AI Agent)

The same style in several colours shares one pattern: upload one garment photo per colour (1–12) plus a pattern (direct upload, or free extraction from a viral photo). Click Start analysis (free) to create a durable task; review the colour list and pattern preview, then click Confirm generation to pay for the plan — the pattern is migrated pixel-faithfully onto every colour and every flat-lay renders automatically. Once a flat-lay is approved, generate its model shot (one model throughout, a different scene for every colour).

Steps

1. Same-style colour garments (1–12 photos): one photo per colour version of the same style; add, remove and reorder freely.
2. Pattern source: choose either Upload pattern (a standalone swatch or a close-up of the print), or Extract from a viral photo: upload a viral garment photo carrying the target pattern and the system immediately starts a free background extraction — the card shows Extracting…, then switches to Extraction ready with a preview once done. Click Re-extract for another attempt, or switch back to Upload pattern to use a different image instead. Re-using the same viral photo hits a cache — instant and free.
3. Model reference photos (optional, 0–2): upload to anchor that exact model's identity throughout; leave empty and AI designs one model matching the garment style and target market, kept the same across every colour.
4. Pattern colour strategy: Preserve original colours (default — pixel-faithful, no recolouring) or Adaptive colour (the motif and layout stay identical; colours are nudged to complement each garment colour).
5. Flat-lay style / model / tier / target market / both ratios: Pattern Transfer supports GPT Image 2.5 Sunburst (default), GPT Image 2.5 Flare, GPT Image 2 HD, Nano Banana Pro, Nano Banana 2, and Seedream 5 Pro, with a separate flat-lay ratio and model shot ratio (both default to 1:1). These settings freeze only when paid generation is confirmed.
6. Start analysis (free): this deliberate click creates the task first, then reads every colour's name (English + local), category, gender, and checks whether every garment photo is genuinely the same style. The task immediately appears in My Creations; leaving the page never interrupts it. No credits are charged at this stage.
7. Review analysis_ready: the colour list (a thumbnail plus editable English/local names for each colour), the pattern preview, and editable category/gender. A garment photo that looks like a different silhouette surfaces a non-blocking consistency notice — continuing is your call.
8. Confirm generation (2 credits for the plan, plus every colour's flat-lay priced by the chosen model — the exact total is shown on the button): this is the first paid action. AI migrates the pattern onto every colour and renders every flat-lay immediately.
9. Results (colour card grid): one card per colour, headed by a colour-name badge, with Flat-lay / Model shot / Pattern vector file below it in generation order — three columns side by side on desktop, the same order stacked vertically on phones. Each card header has a checkbox — checkable once that colour's own flat-lay succeeded, checked by default. Once a colour's flat-lay is a terminal success with an archived thumbnail, the model-shot column's Generate model shot button unlocks — it renders that colour's own scene and pose with the same model identity, priced independently.
10. Multi-select batch generation: once every flat-lay has finished (success or fail) and at least one succeeded, a toolbar appears at the top of the results — a select-all checkbox (toggles between select-all and clear-all) plus two independent buttons, Generate selected model shots and Generate selected pattern vectors, each priced live from the colours that are both checked and don't already have that render. Uncheck a colour to leave it out of this batch; a colour that already has a model shot or pattern vector is never re-charged or re-rendered by either button, even while still checked.
11. Pattern vector file (print-ready, the third result column): once a colour's flat-lay renders successfully, that column shows a Generate pattern vector button (8 credits, regardless of whether a model shot exists) — AI extracts only the printed graphic from that flat-lay and redraws it as a standalone file: white background made transparent, roughly 4096px wide (about 34.7cm at 300dpi, inside the 32-36cm print-size range). The preview renders over a checkerboard backdrop so the transparency is visible; Download saves a PNG with a -300dpi filename suffix. A failed render can be retried in place at the quoted price.
12. Unified edit and regeneration: use Edit & regenerate at the top-right of a card's prompt area to edit the flat-lay prompt, scene and pose together. Changing only scene/pose regenerates only the model shot. Changing the flat-lay prompt first regenerates the flat-lay, then automatically renders a new model shot from it.
13. Other per-card actions: Smart edit (chapter 7 — annotate and overwrite in place), zoom, download and Save to library (a free independent copy that survives task deletion; with folders you pick which one to file it under).

13. Viral Video Remake (AI Agent)

Upload a viral reference video. AI analyses its camera work, pacing and voice-over, swaps in your own character / product / scene, re-renders it segment by segment, and stitches the final cut with one click. Eligible test accounts can also compare the stable Original-video path with an experimental Depth-reference path.

Steps

1. Reference video: MP4/MOV, 3–60 seconds, up to 100 MB. After upload the system detects the voice-over for free and transcribes it word for word. The same analysis always creates a publish title plus 3–5 directly copyable #tags. Pasting the original title and tags is optional and only helps match their structure and tone.
2. Reference mode (controlled test accounts only): Original video · Stable is the default and gives the video model the trimmed source clip. Depth reference · Experimental converts each segment into a depth-motion video before rendering, preserving body/object geometry and motion while suppressing the source clip's colour, texture and identity. The selector is hidden when the experiment is not enabled for your account; saved historical tasks still show the mode they used.
3. Replacement images: a product image is required; character / product / scene share 8 slots. Character photos receive a free automatic white-background pass (“White ✓”). In Original mode an empty group can retain or infer an equivalent from the source, and an empty scene group may use the cleanest source frame. In Depth mode source appearance is deliberately unavailable: only your uploaded replacement images anchor identity, product and scene, so upload every appearance you need.
4. Selling points and unified requirements: add factual product selling points (required; AI fill is available) and optional whole-video requirements for camera, styling or messaging. These inputs are shared by every segment.
5. Voice-over — pick one: AI rewrite writes a new script from the visible product facts and your selling points; Keep original (editable) reuses the detected transcript after your edits. Keeping an empty transcript preserves the reference audio; rewrite can add narration to a silent reference. Subtitles are off by default, and the language picker covers English, Chinese, Spanish, Japanese, Malay and Thai.
6. Choose video-generation parameters: before pressing Analyse, choose the model, resolution and ratio. The whole Seedance family's generation route is scheduled automatically by the platform; Seedance 2.5 appears right after Seedance Mini and offers 480p / 720p only. This complete parameter set freezes when analysis is submitted, and later top-ups and re-renders reuse it.
7. Analyse shots (2 credits): the AI returns a segment-by-segment script — each 4–15 seconds with camera movement, action and its voice-over line. The same analysis call writes directly for the final model, with no extra request or fee: Seedance uses its media-role syntax. Voice-over lines are individually editable.
8. Render: select the segments to render (or select all), then click Generate. Segments render as a sequential chain: the previous segment's final frame is supplied to the next one.
9. Compose the full video: once every segment succeeds, Compose full video appears (free) and stitches the cut. Re-rendering a segment lets you recompose.

Staying in control

14. Video Upscale (AI Agent)

Turn soft, low-resolution footage into a sharp 720p / 1080p / 2K / 4K version — an AI super-resolution operator rebuilds detail frame by frame. The empty workspace shows a real before/after demo: drag the divider to compare.

Steps

1. Pick the upscale model (one of two): Volcengine Video Upscale — choose a target resolution from 720p to 4K, any source size; or Aliyun Video Upscale — a fixed 2× upscale with quality enhancement, requiring a source above 360×360 and below 1920×1080, up to 1 GB, priced by duration at a flat 2 credits/second.
2. Upload the source video: up to 600 seconds (10 minutes); the library works here too.
3. Pick the target resolution (Volcengine mode): 720p / 1080p / 2K / 4K. Portrait footage is measured on its short edge (portrait 1080p outputs 1080×1920). Aliyun mode skips this step — output is always 2× the source.
4. Check the quote: the button shows the exact price — it scales with duration, frame rate and target resolution — alongside an estimated processing time. Duration and frame rate are probed server-side, which is what the charge is based on.
5. Click Upscale. The task runs in the background with a live “time remaining” estimate; you can leave the page.
6. When it finishes, the result appears on the right — click Download to save it.

15. Motion Transfer (AI Agent)

Transfer the motion from a reference video onto your own character photo: upload one character photo plus one motion video, and AI makes the subject in the photo perform the motion from the video. The result runs as long as the motion video and keeps its audio.

Choose a model first

The page offers two complementary models; the choice changes both the available parameters and the price:

In short: pick Kling when you need 1080p, otherwise pick Wan — it is cheaper, takes longer videos, and is the only one that handles non-human subjects.

Steps

1. Upload the character photo (1 image): it must clearly show the subject's head and torso. JPG / PNG, up to 10MB, min 340px on the short side, aspect ratio between 2:5 and 5:2.
2. Upload the motion reference video (1 clip): MP4 / MOV, up to 100MB, min 340px, and its length must fall inside the selected model's range (see the table above).
3. Transfer mode (Wan only): Into the video (default) swaps the subject in the video for your character and keeps the video's own scene and camera; Onto the photo applies the motion to your character photo and keeps that photo's own background.
4. Match the video's scene (Kling, and Wan's "Animate the photo"): before submitting, the system composites both the setting and the subject's pose from the motion video's first frame into your character photo for free — the result then shares the video's background and starts from a matching pose, so the motion picks up naturally. Turn it off to keep the photo's own background and pose. Compositing takes about a minute, during which Generate stays disabled. Once it finishes, the character photo is replaced by the composite, so you can confirm exactly what will be rendered before submitting. If you don't like the result, add a line of extra direction in the box below (for example "stand closer to the camera" or "no glasses") and press Regenerate the scene image — the extra direction applies only when you press that button. Wan's "Into the video" already uses the video's setting natively, so the switch is hidden there.
5. Pick a resolution: follows the model (Wan 480p/720p, Kling 720p/1080p).
6. Pick the character orientation (Kling only): "Match video" reproduces complex motion more faithfully; "Match photo" supports more camera movement.
7. Prompt (optional): can be left empty. Add one only when you want to emphasise a specific detail of the motion.
8. Click Generate. The button shows the exact credit cost.

Pricing

Both models are charged by the length of the motion video (the result runs exactly as long; partial seconds are not counted). Wan 2.2 Animate: 8 credits/second at 480p and 16 at 720p, billed as at least 5 seconds. Kling 3.0 Motion Control: 20 credits/second at 720p and 27 at 1080p. An 8-second motion video at 720p therefore costs 128 credits on Wan and 160 on Kling. The server re-reads the video's real duration before charging, so the final amount always follows the true length.

16. Video Edit (AI Agent)

Change a finished video with one sentence: modify / add / remove visual elements, edit the sound, extend it, or splice 2–3 clips with seamless transitions. Upload a video directly, or jump in from the “Video Edit” button on any AI result card.

Six task types

Selecting a task prefills the officially recommended sentence template — just replace the angle-bracket placeholders. Text you have edited yourself is never overwritten when switching tasks.

Steps

1. Add the source video: upload, pick from the library, or bring in a result. Per-clip length caps follow the model: 30s on Seedance 2.5, 15s on the 2.0 family; splice accepts 2–3 clips.
2. Browse the filmstrip: a CapCut-style timeline under the player (evenly sampled to fill the width, no scrolling) — tap or drag to move the playhead.
3. Scope the edit range (Seedance 2.5): press and drag directly across the strip to select the range (the outside dims; edge handles fine-tune, a single click still seeks); “within Ns–Ms” is written into the instruction automatically so the edit only touches that window. Drag back to full width to clear it.
4. Pick a task and finish the instruction: replace the placeholders, or write your own.
5. (Optional) auxiliary reference images: for combo tasks like swapping products or prints, attach a product photo and reference it as “referencing Image 1’s product…”.
6. Choose model and resolution: all four Seedance models. Seedance 2.5 edit tasks automatically match the source’s length and aspect ratio (locked by the official spec); the 2.0 family and extend tasks set an explicit output length.
7. Click Generate (the button quotes the price live). When done, use the Original / Edited toggle on the player for instant A/B comparison; continue editing the result for chained iterations, or send it to Video Upscale.

Pricing

The existing reference-video rate applies: (output seconds + floored source seconds) × credits per second. Example: editing a 13s source on Seedance 2.5 at 480p = 17 × (13+13) = 442 credits; the same task on 2.0 Mini at 480p with a 5s output = 2.4 × (5+13) ≈ 43.2 credits. Splice output is the sum of the clips plus ≈1s per transition. Failures refund automatically.

17. Text to Speech (AI Agent)

Turn a script into natural spoken audio: type or paste the copy, pick a voice and a model tier, and synthesize an mp3 in one click. Pause markers, emotion tags (Expressive tier) and AI script polish are built in; history and examples live on the right side of the workspace.

Steps

1. Enter the script: up to 2,000 characters (billing counts characters, inserted markers included). The toolbar inserts pause markers — the menu follows the tier: the Standard tier offers exact durations (0.3 / 0.5 / 0.8 / 1 / 1.5 / 2 / 3 s, read as silence of that length), while the Expressive tier offers short / medium / long semantic pauses (about 0.5 / 1.5 / 2.3 s) because that model does not support timed pause markers; Smart markup (2 credits) analyzes the script and inserts fitting pause markers and emotion tags (pauses only on the Standard tier) without changing a single word.
2. Pick a voice: "Change voice" opens the catalog — 16 curated female/male voices with gender filter, name search, previews and favorites; the "Custom voices" tab lists the voices you cloned (see the "Voice Clone" section below). Your voice is remembered — the workspace preselects the one you used last time (falling back to the built-in default if that voice was deleted).
3. Pick the model tier: Standard (3 credits / 100 chars) is fast and economical with full speed / stability / similarity / style sliders; Expressive (5 credits / 100 chars) delivers the most natural read and unlocks emotion tags — the toolbar groups 50 common tags under Laughter / Emotion / Breath & mouth / Delivery / Other (e.g. [laughs], [whispers], [professional], [warmly]); the model accepts open-ended descriptors, so you can also type your own, such as [frustrated sigh]. Stability comes as a three-preset control (Creative / Natural / Robust).
4. Click Generate speech (the button shows the live price; under 100 characters bills as 100). The task appears under History and refreshes to completion automatically.
5. Play the result right on the card, Download the mp3, Reuse script to edit again, or Delete (which also destroys the result file).

Voice Clone

Upload your own reference audio and clone a personal voice — once ready it appears in the Text to Speech voice catalog under "Custom voices", and every history card offers one-click "Use in Text to Speech".

1. Name the voice and upload 1–3 reference recordings (30 s–3 min of clean speech each works best — no background music or noise; the asset library works too).
2. Click Clone voice (100 credits each). Cloning usually finishes within a minute, and the finished card carries an auto-generated preview.
3. Hit "Use in Text to Speech" to jump straight into synthesis with that voice — or pick and favorite it any time from the voice catalog's "Custom voices" tab.

18. Viral Video Insights (Ops Agent)

The first ops agent: upload a benchmark viral video with its original link. The model watches the footage and listens to the audio, then breaks it down into 31 dimensions across 7 modules and keeps the result automatically as the final version in a library shared by your team (the master account and its members) — accumulated over time to guide AI video generation later.

Steps

1. Upload the video: a single downloaded benchmark clip (≤180 s). It is used for this breakdown only — the copy is destroyed as soon as the analysis finishes and only the link and metadata are kept.
2. Enter a tag: required — a label that groups this breakdown, such as "Top 100 apparel" or "trending now"; keep one spelling per tag. Tags your team used recently are listed under the field for one-tap fill, and your last value is pre-filled. Analytics filter by tag.
3. Paste the original link: it records the source and identifies the platform (TikTok / Douyin / Instagram / YouTube / Xiaohongshu / Kuaishou). Views, likes, comments, price and other platform data are not entered in this phase; the fields are reserved and show "Pending API".
4. Press Start breakdown; it usually finishes in 20–60 seconds and the page refreshes itself.
5. Read the breakdown: the detail page expands 31 dimensions under 7 modules (source, script & structure, garment selling points, model & on-body proof, native feel & audiovisual, comment feedback, reuse & shooting conclusions). Every dimension carries an observation, evidence timestamps and one of four labels — Fact (verifiable on screen or in the audio), Inferred, Not shown, N/A. A finished breakdown is kept as the final version automatically (no manual confirmation since 2026-09-09); the 31 dimensions are read-only, and changing them means a new upload.
6. Check the coding: the "Coding" card on the detail page shows the 12 fields the model assigned from the codebook (opening hook, shooting style, editing, audio mode …). They can still be corrected after the breakdown is kept; corrected fields carry a "Corrected" badge. These values are what the Analytics page aggregates, and a correction counts within about a minute.
7. The library: the library is scoped per team: a master account and its members share one library, and an account without a team sees only its own records. A record stays with the team it was created in — leaving a team does not take it along and joining one does not bring it in. Only the submitter, the team master or an admin can delete a record. The Analytics page's "library" likewise covers your team only. The list filters by tag (including "Untagged") and every row carries a tag badge; records from before this field existed have no tag — tick several rows (or select the whole page) and enter the tag once in the bulk-tag bar, or edit a single record on its detail page.

19. Viral Insights Analytics (Ops Agent)

See the patterns across every confirmed breakdown: how viral videos open, allocate time, which shooting / editing / sound / caption choices they make, which selling points they show and how the body proves them, which technique combinations keep recurring, and how reliable the evidence in the library is. Only confirmed breakdowns are counted, and every chart can be compared against the whole-library baseline after filtering.

The codebook underneath

Breakdown observations are free text; counting them directly only yields fact / not-shown rates and tag frequencies. So after every breakdown the system runs a text classification pass that maps 12 key fields onto a codebook: category, opening hook, ending CTA, shooting, editing, audio mode, captions, scene, angles (multi), comparison, selling methods (multi) and reusable elements (multi). Model codes take effect by default; operators only correct mistakes in the "Codes" card on the detail page — while a breakdown is under review the dropdowns and chips are editable, edited fields carry an "Edited" badge, and "Recode" reruns the model without ever overwriting a human correction. When the codebook changes, the analytics footer shows how many confirmed rows still use the old version and an admin can recode them in one click.

What the page shows

1. Entry points: sidebar "Ops Agents → Analytics", or the "Analytics" button in the breakdown library header. It shares the module switch with Viral Video Insights.
2. Filter bar: confirmed time (all / last 30 / last 90 days), platform, category, duration bucket (≤12 s / 12–25 s / >25 s), tag (including "Untagged" — once picked, every chart and KPI counts only breakdowns under that tag). With "Show library baseline" on, the short tick on each bar is the whole-library share, so a category's difference from the whole is visible at a glance. The right side reads "Filtered n / library N".
3. KPI row: confirmed count, new in 30 days, median duration, median segments, voice-over / captions / one-take shares, platform-data progress (pending API).
4. Structure: opening hook, hook length, opening / middle / ending time allocation, ending CTA.
5. Audiovisual: shooting × editing heat matrix, audio mode, captions, scene.
6. Selling points & on-body proof: how often each of the six selling-point dimensions is shown (with a callout for how many viral videos never address sizing), angle coverage, comparison type.
7. Combinations: pairs and triples across opening hook / shooting / editing / audio / captions that keep appearing together, with support and counts.
8. Style tags, evidence quality, trend: normalized top tags and co-occurrence; fact rate per module and the dimensions most often "not shown" (the gaps a reshoot can fill); confirmations per week and key shares by month.
9. Recent reshoot plans: the reshoot plans of the five latest confirmed breakdowns, ready for the shooting team.
10. Export CSV: the top-right button exports the confirmed rows under the current filter, one per row, with the 12 codes, the 31 evidence labels, normalized tags, source link and summary.

20. Listing image insights (ops agent)

The second ops agent: batch-upload a benchmark product's Amazon listing and A+ images by folder. The model reads every image for composition, lighting, subject presentation, typography, copy and color, then looks at the whole set for structure, narrative order, visual system and selling-point coverage; each set is kept in your team's library and coded automatically the moment it finishes, ready for "Image analytics".

Pick the set type, then upload folders, zip files or images

1. Pick the set type: click "Listing set" or "A+ set" before anything else — one type per batch, and the three upload entries stay disabled until you do. A Listing set is the main and secondary images (images whose folder or file name carries an A+ hint are still recorded as A+); an A+ set is A+ page modules only, with no main image, and is analyzed and coded as an A+ page.
2. Prepare and upload (any of three ways, mixable): (1) several folders — each folder holds one set's images directly (main.jpg or a MAIN file name marks the main image, the rest sort as 1.jpg, 2.jpg…; the ASIN-product-images.zip and ASIN-A+-images.zip a scraping tool exports work too); drag them together onto the drop zone, or click "Choose folders" and pick their parent directory — the page groups by sub-folder (the browser's directory dialog picks one folder at a time, so the parent directory is the multi-set entry); (2) several zip files — "Choose zip files" takes many zips at once, each zip is one set, titled by the ASIN in the zip name with an Amazon link; (3) several images — "Choose images" collects loose files into one set. A leading index in a folder name (such as "3 B00D1ARZMC") is dropped. Within a set the main image comes first and the rest follow file-name order; no per-set image limit (sets over 10 images are analysed in batches of 10 and each batch is saved as it finishes; a 200-image safety ceiling only guards against mistakes); up to 20 sets per batch. The page previews the groups first, with the image and A+ counts of each.
3. Analysis progress: images are analysed 10 at a time and every finished batch is written to the record — the list row shows "x / y images analysed", the detail page shows the per-image breakdowns of the finished batches, and the set-level dimensions, summary and style tags are compiled once every image is done. If a batch fails, the finished batches are kept and Retry analysis on the detail page resumes from there without re-running finished images.
4. Enter a tag: required, one per batch (e.g. "Top 100 apparel", "trending now"); keep one spelling per tag. Recently used tags are listed under the field and your last value is pre-filled. A set that belongs under another tag can be changed on its detail page.
5. Start: press "Start breakdown (N sets · M images)". Images are resized to 1600px in the browser before upload (4 at a time); each set starts analyzing as soon as its upload lands, and the list header shows how many sets of this batch are done or failed. A failed upload voids only that set.
6. Review: open a record from the list: an image strip at the top (click to enlarge), then the summary, the codes card, the 6 set-level dimensions and the 9 per-image dimensions (first impression, style reference, presentation, composition, lighting, background & palette, main purpose, typography, copy — redefined on 2026-09-12 after a design-direction table). Every observation carries a fact / inferred / not shown / n/a tag; the copy dimension transcribes on-image text verbatim, and typefaces are classified by style only (bold sans, serif, script…) with font names as guesses at most.

21. Image analytics (ops agent)

See the patterns across every kept image-set breakdown: how main images are composed and what background they use, how lighting and color are handled, how type and copy are laid out, how the product is shown, in what order sets tell their story, which combinations keep recurring and how reliable the library's evidence is. Only kept sets count, and every chart compares the filtered slice with your team's whole library.

The codebook

Each finished set is classified once into 12 enumerated fields: category, main-image composition, main-image background, lighting style, typography style, text density, copy focus (multi), subject presentation (multi), color scheme, set structure, A+ modules (multi) and reusable elements (multi). Model codes apply by default; operators only fix mistakes on the detail page's codes card, and corrected fields carry a "corrected" badge. "Recode" re-runs the model values and never overwrites manual corrections. After a codebook upgrade the analytics footer shows how many sets use an older version, and an admin can recode them in one click.

What the page shows

1. Entry points: sidebar "Ops agents → Image analytics", or the "Image analytics" button on the library page. Same module switch as listing image insights.
2. Filters: kept date (all / last 30 / last 90 days), category, type (Listing / A+ sets — main-image composition, background and pure-white share count Listing sets only), tag (including "Untagged" — once picked, only sets under that tag are counted); with "Show library baseline" on, the short tick on every bar is the whole-library share. The right side reads "Filtered n / library N sets".
3. KPI row: kept sets (with the last-30-day count), median images per set, and the shares with A+, with on-image text, with a model wearing the product and with a pure-white main image, each against the baseline.
4. Main image: composition, background, and a composition × lighting heat map.
5. Lighting and color, text and copy, subject presentation: lighting style and color scheme; typography style, text density and copy focus; subject presentation and category.
6. Set structure: narrative structure, images per set, A+ module types.
7. Combinations, style tags, evidence quality, trends: recurring combinations of composition / background / lighting / typography / color; normalized tags and co-occurrence; fact rate per dimension and the most often "not shown" dimensions; sets kept per week and the model-wearing / on-image-text shares by month.
8. Recent takeaways: what the latest five kept sets are worth borrowing and where they fall short, for the design team.
9. Export CSV: the current filter's kept sets, one row per set, with the 12 codes, per-dimension fact rates, normalized tags and the summary.

22. Ops Board (ops agent)

Per category line (e.g. Vecardi, Viakeo) the board pulls TikTok Shop data from EchoTik every day to answer one question first: who is selling this category, and are they worth contacting. The first board is the Creator board: creators are found backwards from the day's product pool, then graded and scored on the golden metrics — category GMV and its deltas, GMV per 1k followers, sales mode, e-mail — and one click on a row opens due-diligence metrics and the outreach record. Product / content / competitor boards are coming next. Free, no credits.

First-time setup

1. New line: sidebar "Ops agents → Ops board", then "New line". One line is one product line: name (e.g. Vecardi), region (US by default), TikTok category (pick down to the third level in the dictionary — the daily rank is only pulled for a level-3 category; Vecardi is Womenswear & Underwear › Women's Special Clothing › Workwear & Uniforms; if the dictionary fails to load, switch to "Enter category id manually"), own shop (search by TikTok Shop name and pick it; linked creators are pulled from it, own products stay out of the product pool), price band (USD, used to check whether a creator's 30-day average selling price fits), keyword seeds (comma-separated, up to 8; every seed is searched daily under four sort orders, 30 products each, until the 100-product pool is full), daily call cap (10–20,000, default 10,000; the product pool takes about 10 calls, product details and shop profiles up to 10 each, then one call per product for its creators and one per creator for the profile and one for videos — a line usually spends a few hundred a day) and the "Enable daily sync" switch.
2. Daily sync: from 20:00 Beijing time the system pulls yesterday's data (EchoTik data is T+1; the US day only ends at 15:00 Beijing time and the category daily rank usually appears in the afternoon, so a daytime manual refresh shows a "not published yet" notice for it) and re-checks blocks that were not yet available at 22:00; the alert strip names blocks by type ("Category daily rank", "Keyword", "Creator profiles", …). The header shows the "Data as of" date, the sync state (done / partial / running / failed) and the usage line "Today x / cap · month y"; skipped or failed blocks are listed in the alert strip.
3. Refresh: each line can be refreshed manually once a day; after a settings change one more refresh is allowed and it re-pulls the affected blocks under the new settings. While a sync runs the page updates every 15 seconds.
4. History: the date picker in the header switches between the last 60 data days; the dot in front of a line shows enabled / paused.

Where the creator pool comes from

• Product pool first: every day each keyword seed is searched under four sort orders (7-day units, 7-day GMV, lifetime GMV, creator count; sort order first, seed second; 30 products each, at most 16 searches a day) until the day's pool holds 100 products after de-duplication — once it is full the remaining searches are not sent; then the category daily rank top 30 is pulled, which only ranks products already in the pool and fills spare slots if any; own products stay out of the pool (the "Product pool" table at the bottom is paginated over every pooled product, 20 / 50 / 100 per page: in the # column "kw" = keyword hit; plus shop, average price, period units and creator count). Changing seeds during the day does not shrink the pool; new seeds take full effect from the next data day.
• Shop info: search and rank rows carry no shop, so the sync also calls the product-detail endpoint to fill each pooled product's shop id and then pulls shop profiles per shop (up to 10 per call, shared site-wide, refreshed every 7 days). Click a shop name in the "Shop" column of the product pool to open its profile: platform label (e.g. Gold Star Seller), rating, positive feedback, response and shipping rates, followers, products listed, 7-day / 30-day / lifetime units and GMV, selling creators and the TikTok shop link.
• Creators looked up per product: for the pool's products the selling creators are pulled one product at a time (the whole pool, at most 100 products; rank products first, then keyword products), and each creator's sales on the pooled products are summed into category GMV / category units — attribution to this category specifically, not the creator's all-category total.
• Profiles and videos: missing or 7-day-stale creator profiles are completed in batches (up to 400 creators a day: followers, 30-day GMV, average selling price, interaction rate, score, e-mail…); every creator of the day, in category-GMV order, gets their top product's selling videos (within the daily call cap), which give the latest video, views and GMV per 1k views.
• Daily rows and deltas: a creator day-row is written every day and compared with 7 and 30 days earlier for the 7-day / 30-day deltas; a first appearance is marked "New"; day-rows are kept for 180 days.

What the board shows

1. KPIs: pooled products, creators, grade-A creators, new in the last 7 days, creators linked to your own shop.
2. Grades: A = sells ≥2 pooled products, 30-day category GMV is growing (a first appearance counts when it has sales) and has a contact e-mail; B = category GMV top 50; C = everyone else. Follower tiers: under 10k / 10k–100k / 100k–1M / over 1M.
3. Creator pool table: grade, creator, category GMV (descending by default; click a header to re-sort), category units, 7-day delta, 30-day delta, products driven, top product, GMV per 1k followers (category GMV ÷ followers × 1000), sales mode (video / live / video + live / showcase), latest video (days ago), e-mail, outreach status; 20 rows per page (switchable to 50 / 100) with the pager under the table, back to page 1 whenever a filter or sort changes. Filters: A / B / C, tier, mode, outreach (not contacted / contacted / an exact status such as contacted, sample sent, published, dropped), e-mail only, active within 60 days, search by name or handle; "Export list" saves the current filter result as CSV (profile link, e-mail, outreach status and note included).
4. Own-shop block: once the shop is matched it shows linked creators (creators selling your products and their own-shop GMV) plus who was added or lost since the previous day; own products stay out of the pool, so creators per own product and partner videos are no longer listed.

Due diligence and outreach

• Category performance: category GMV · units, GMV per 1k followers, 7-day / 30-day deltas and every product driven (GMV · units · price).
• Due-diligence metrics: 30-day total GMV with the video share, 30-day average selling price and whether it fits the line's price band (a deviation over 40% is flagged), selling videos' 7-day average views with the interaction rate, 30 / 90-day follower growth, platform score, lifetime GMV and products sold in 30 days; "Profile pending" shows until the profile is completed, and a likely-deactivated account is flagged.
• 90-day trend: "Load 90-day trend" pulls daily GMV and daily follower gains on demand (about 9 calls, cached for a day).
• Pooled-product videos: the creator's videos for pooled products — views, units, GMV, GMV per 1k views, publish time and length — each opening in a new tab.
• Outreach record: status (candidate / contacted / sampled / published / dropped) + a note (channel, quote, sample status, ≤500 characters) + the published video id; "Save outreach" writes the status into the table's outreach column, shared with the whole team.

23. Batch Video Remake (Batch Production)

"Batch Production" is a new top-level group beside AI Video, AI Image and AI Agents. Its first module is Batch Video Remake: one viral reference video + character photos + a scene photo + the product images of a batch of same-style SKUs; press "Generate batch" once and every SKU gets its own video with the same shots and the right product — nothing to do after launch.

Steps

1. Reference video: upload one viral clip (or pick it from the library). Every SKU video mirrors its action, camera work and pacing shot by shot. When the source is longer than the model's reference cap (15s for the Seedance 2.0 family, 30s for 2.5) a "Reference span" block appears: drag the "Start" and "Length" sliders to pick the span to replicate; the output length follows that span.
2. Character and scene photos: 1–2 character photos (one clear front-facing face per person — face, hair and build come from here; do not upload several angles of the same person); an optional scene photo — when given, the environment is rebuilt from it, otherwise the reference video's environment is kept.
3. SKU groups: one card per SKU with a label (e.g. "Navy", "Wine") and 1–5 product images — the first is the main image (silhouette and colours), the rest are detail shots (fabric, seams, hardware). "Add SKU" adds a card, up to 20; a batch must be different SKUs of the same style, and the first SKU produces the master prompt.
4. Parameters: model (Seedance 2.0 / Fast / Mini / 2.5), resolution, aspect ratio, audio handling (follow the source rhythm / music only / voice-over / no constraint); optional product facts (construction, selling points, audience) and extra requirements every prompt must follow.
5. Voice-over (added 2026-09-13): picking it reveals the VO language (six options, defaulting to the site language) and a Subtitles toggle. After submission the AI writes one voice-over from the reference video's shot pacing and your product facts (the same writer the prompt wizard uses); each SKU's script is rewritten from its own product facts (when they match the master's, the master script is reused with only the product wording swapped), keeping the master's time segments and pacing. If no usable script can be written the whole batch fails and the analysis fee is refunded — it never silently ships a clip without the voice-over. Filling in the product facts and selling points makes the script noticeably better.
6. Per-SKU product facts (added 2026-09-13): the product info card is shared by the batch and whatever you type wins for every SKU; style and selling points you leave empty are read by the AI from each SKU's own product images (only what is visible), while the audience stays one per batch. If that read fails the SKU falls back to your typed facts only — it never reuses the first SKU's inferred facts.
7. Quote and launch: the quote block shows "Analysis 2 + SKU count × (plan 2 + current render price) = total" plus the output / reference seconds; press "Generate batch · N credits". Only the analysis fee is charged at submission; each SKU's plan and render fees are charged when its turn comes.

How the batch is replicated

• One analysis for the whole batch: the reference video is broken down into the goal, shot-by-shot visuals, camera work and environment — the same logic as the prompt wizard's "Analyze the video" step in Reference to Video, without the per-section confirmations.
• Master prompt + per-SKU rewrite: the first SKU's product images produce a master prompt that is lint-checked and patched automatically; every other SKU only rewrites the product description on top of it (colours, materials, signature details and the referenced image numbers), and a rewrite that does not pass falls back to composing that SKU on its own — same shot plan, correct product each time.
• One render per SKU: each SKU is an independent reference-to-video task rendered on the chosen model and archived automatically.
• Identity lock and environment-free actions: the actions broken down from the reference video describe only the person's movement and orientation, and the environment comes only from the scene photo and scene text; after linting, every SKU prompt gets a mechanical "Identity lock · top priority" block (one person throughout, every angle follows the character photo, back shots show that hairstyle as seen from behind, people in the reference video and product images are only motion sources or mannequins), restated on every turning shot. The block is visible in "View prompt". For steadier identity, add a side-back or back view of the character and prefer flat-lay or white-background product images.

Review, stop and re-render

• Grid cards: each SKU shows queued / writing prompt / ready to render / rendering / completed / failed; the first SKU carries a "Master" badge. A finished card plays and downloads, and "View prompt" shows the exact prompt used for that SKU and where it came from (master / rewritten from master / composed separately).
• Stop: while a batch runs, "Stop" in the top-right corner (with confirmation) refunds the plan fee of SKUs not started, stops and refunds renders in flight, and keeps finished videos.
• Re-render: a failed SKU card offers "Re-render · N credits" to render that one again at the current price.
• Adjust & retry: in view mode, "Adjust & retry" loads this submission's materials, span, SKUs and parameters into a new form so you can change them and launch another batch.
• My Creations and billing: My Creations shows one card per batch (type filter "Batch Video Remake") with the SKU count and the net credits of the whole batch; billing folds the batch into one "Batch Video Remake" entry that expands into the analysis fee, every SKU's plan / render fees and refunds.

24. My Creations

All of your work in one place — videos, images, storyboards, listing sets, replicas, garment pattern variations, pattern transfers, remakes and upscales — as cards you can filter, open, download and delete.

Filtering

• Task type: a row of buttons across the top (All / Reference to Video / Image to Video / Text to Video / Image to Image / Text to Image / Storyboard / Product Listing Set / Viral Video Remake / Image Replica / Garment Pattern Variations / Pattern Transfer / Video Upscale / Batch Video Remake) — click to switch.
• Status + time: the second row filters by status (generating / done / failed) and by time (today / this month / custom range).
• Filters are remembered: your chosen type, status and time range are restored automatically the next time you open the page.

Card contents

• Thumbnail and status badge: generating / done / failed at a glance; video results carry a play icon.
• Prompt, model, parameters and task ID below the thumbnail; agent tasks also show render progress and a strip of source material.
• Credits used: the total for that task (agent tasks show the net total for the whole pipeline).

Common actions

1. Open / retry: click the card body to open that workspace in view mode; Adjust & retry there loads the parameters into a new task.
2. Download: save the result with the Download button.
3. Delete: the trash icon in the bottom-right corner of the card, with a confirmation dialog. Deleting removes the work and its result files, and refunds anything still running.

25. Asset library

Materials is one searchable library for your uploaded images, videos and audio. One-level folders keep a large library manageable without making extra file copies.

18.1 Views, search and folders

• All assets shows the complete library. Recent is ordered by assets you used most recently, Favorites keeps marked assets together, and Unfiled contains assets that have no folder.
• Folder names appear in the Materials page's own rail on desktop and in the view selector on mobile. Create, rename or delete one-level folders from the same navigation.
• Use the search field to find a filename or folder. Type, reference-status and sort controls narrow the results; pages are loaded from the server, so older assets remain discoverable as the library grows.

18.2 Organize without copying

1. Select one or more rows, then Download selected, Move to folder, Move to Unfiled, favorite or delete them in one batch.
2. Download selected always delivers the original files: a single selection downloads the file itself; several are packed into one zip (original upload names, duplicates suffixed -2 / -3) with no trimming, transcoding or recompression — videos longer than 15 s come down whole. One archive is capped at 500 MB in total, so split larger selections; any asset that cannot be fetched is skipped and named in the notice.
3. Each asset has at most one primary folder. Moving it only changes organization metadata; the underlying file remains stored once.
4. Deleting a folder never deletes its materials. They remain in All assets and move to Unfiled.

18.3 Use folders from a workspace

• Uploads made inside video, image and agent workspaces enter Unfiled by default. Organize them later from Materials.
• The Library picker (Figure 6-4) can filter All assets by folder. Picking records Recent use only and never moves the source asset.

18.4 Deletion rules

1. Use a row action or the batch bar to delete the underlying asset. Batch deletion reports deleted and skipped rows separately.
2. If an asset is still referenced by a task, deletion is blocked and the protected row is skipped. Moving or favoriting it is still safe.

26. Account settings

The Settings group in the left menu holds your profile, security and billing pages.

26.1 Profile

1. Open Settings → Profile.
2. Update your name, avatar and other details, then save.

26.2 Security & phone binding

Phone binding remains available as an account-security capability; the product sign-in page itself offers email/password and Google:

1. Open Settings → Security and find the Phone card.
2. Enter your 11-digit number and click Send code.
3. Enter the code and click Bind. The card then shows the masked number.

The same page lets you change your password — enter the current and new password, then save.

26.3 Credits & billing

• Balance: your current balance plus lifetime earned and spent totals.
• History: sign-up bonuses, task charges/refunds, top-ups and adjustments with their time and linked task; related task movements are grouped for readability.
• Grouped agent tasks: the sub-charges of one storyboard, listing set, replica, garment pattern variation, pattern transfer or remake are merged into a single row (“Storyboard · 3 entries”); Details expands them and View task opens the work.
• Export your usage: click Export details on the Credit history card and choose the last 7 / 30 / 90 days, a custom range (up to 366 days), or all records. The UTF-8 CSV keeps one row per logical task and lists total deducted, total refunded, net consumed, balance movement and the balance after the latest movement; sign-up bonuses, top-ups and manual adjustments remain separate rows. An all-time export includes the newest 10,000 grouped records and shows a warning if older records were omitted.
• Current plan: shows your plan (Free). Online credit purchase is coming in a later phase.

27. FAQ

Q1　It still says “Generating” — how long should I wait?

Images usually take tens of seconds to a few minutes and videos a few minutes. The limits are 5 minutes for images (10 for GPT Image 2 Flex), about 17 minutes for videos, and up to several hours for video upscaling depending on the job (the UI shows a live estimate). Anything past its limit fails and is refunded. You can leave the page while a task runs.

Q2　My task failed — do I get the credits back?

Yes. Tasks that fail or time out are refunded in full automatically, and the refund appears in your billing history. A failed smart edit leaves the current image untouched.

Q3　I'm not receiving the phone-binding SMS code.

Make sure you are using a mainland-China (+86) number and that international SMS is not blocked. Retry later or contact support if it persists; you can still sign in directly with your email and password.

Q4　Can I recover the original after a smart edit?

Yes. Edits never overwrite the original: the version history under the preview keeps the original and every edit, and Set as current switches back for free. Edits completed before the version history shipped (2026-09-15) did not keep their originals.

Q5　Why do some older images have no “Smart edit” button?

Smart edit needs a high-resolution working copy of the image. Very early results predate that copy, so the button is hidden for them. Everything generated since then supports it.

Q6　Text inside my images comes out wrong in Chinese.

Image models are unreliable with non-Latin scripts, so text rendered inside images is always English in the storyboard and agent flows. Spoken voice-over and subtitles can be in any supported language.

Q7　I've run out of credits.

Clicking a paid generation action shows an “Insufficient generation credits” message and does not create a task or charge your account. Top up from the Pricing page (top navbar or sidebar): four one-time tiers, currently paid by corporate bank transfer via the contact shown in the dialog. The top-up shows in your billing history.

Q8　How is Viral Video Remake charged, and can I stop midway?

Shot analysis costs 2 credits; after that each segment is charged separately (by model, resolution and duration) and failed segments are refunded automatically; composing the final cut is free. You can hit Stop generating at any time — unfinished segments are refunded, finished ones are kept and can be topped up later.

Q9　Are “planning” and “rendering” charged separately in Product Listing Set / Image Replica?

Yes. Planning (the AI writing the set plan and prompts) costs only 2 credits. Review the plan, then batch-render — each image is charged independently by model. If you don't like the plan, Adjust & retry starts over; the small planning fee is not refunded.

Q10　What does Video Upscale cost and how long does it take?

The price scales with duration, frame rate and target resolution (720p is the cheapest tier, 4K about 11× that). The exact quote and an estimated processing time are shown before you submit. Long footage at a high resolution can take hours — you can leave the page — and failures or timeouts are refunded automatically.

Q11　Why does the Garment Pattern Variations workflow render the flat-lay before the model shot?

Upload your images and AI freely auto-identifies each one's role plus an overall style; confirm or edit it and submit once, and the system drafts the plan and renders every flat-lay automatically (2 credits for the plan plus every flat-lay priced by the chosen model — the total is shown right on the submit button), with no separate plan-review or per-card render-selection step. The place you still approve before spending more is the model shot: a card's Generate model shot button only appears once that flat-lay has rendered successfully, so you never pay for a model shot on a pattern you don't like — once every flat-lay is done, a batch toolbar appears at the top of the results, pre-checking every card whose flat-lay succeeded; uncheck any card you don't want, then click Generate selected model shots or Generate selected pattern vectors, each priced live from the checked-and-not-yet-rendered set, never re-charging a card that already has that render. For later changes, use Edit & regenerate in the prompt area: model-only edits charge one model shot, while a flat-lay prompt edit clearly prices and durably runs the flat-lay then model-shot chain, keeping the model on the latest flat-lay. The whole task (plan + every flat-lay + every model shot) is grouped into one "Garment Pattern Variations" billing entry.

Q12　Does pattern extraction in Pattern Transfer cost credits? How is it different from Garment Pattern Variations?

Pattern extraction is always free: upload a viral garment photo and the system starts a free background extraction immediately; the result is content-addressed to the source photo, so re-extracting the same photo again is instant and costs nothing. A failed extraction can be retried with Re-extract, or you can switch to Upload pattern instead. The difference from Garment Pattern Variations is the input shape and the goal: Garment Pattern Variations turns "1 garment photo + a pattern reference" into 10–20 different pattern variants, while Pattern Transfer turns "1 pattern + 1–12 existing colour versions" into that SAME pattern migrated pixel-faithfully onto every colour — better suited to when you already know the exact pattern and just want to cover every colour you sell it in. Both share the same flat-lay-then-model-shot two-stage billing and the "approve the flat-lay before spending on the model shot" gate, and are tracked and billed as separate entries in My Creations and on your ledger.

24. Batch Image Swap (Batch Production)

The second module in "Batch Production": upload a finished set of shots plus the product photos of your other SKUs, press once, and every SKU gets the same set — same person, pose, framing, background, layout and on-image text, with only the product swapped.

24.1 How to use it

1. Reference shots: upload the finished images of one SKU (up to 12), or pick them straight from the asset library.
2. SKU product photos: one card per SKU (up to 10), 1-3 photos each — the first sets colour and pattern, the rest are detail views of that same product. A colour or variant name per SKU goes into the prompt.
3. Choose the model and tier (Nano Banana 2 · 2K by default) — GPT Image 2 / 2.5, Nano Banana Pro / 2 and Seedream 5 Pro — and add one line of extra requirements if needed ("keep the buttons silver").
4. Press Swap batch. The quote before you submit is shots × SKUs = total images. Nothing to do after launch.

24.2 Pricing

• Prompts are built mechanically with no AI planning step, so there is no analysis or plan fee — only the per-image render price.
• Charged cell by cell, right before each render, and refunded per cell on failure. Cells that never start are never charged, so stopping a batch or running out of credits costs nothing extra.

24.3 The result matrix

• Results are grouped by SKU, each group holding that SKU's full set. Any single cell can be zoomed, downloaded, saved to the library or re-rendered at the current price.
• Download all packs the batch into one zip with a folder per SKU, keeping the order and names of the reference shots.
• Color lock is on by default: after each cell renders, the untouched background, people and text are restored to the reference shot's original pixels while the product keeps its new render. Because the reference shows the old product, a swap is never shifted toward the reference's colours, and a SKU close in colour to the old one is never pasted back. The page shows how many are still queued; the work runs in your browser and resumes when you reopen the page.

24.4 Product lock: the product follows your photos, the reference only sets the layout

Each SKU's product photos are that SKU's product lock. The two sides have strictly separate jobs:

• The lock decides: the colours and their proportions, the pattern (motif, scale, direction, repeat and the colours inside it), and the structural parts and hardware with their placement, such as collar, placket, buttons, pockets, piping, cuffs, hems, waistband and drawstring. The colour name you give a SKU is only a hint; where it disagrees with the product photos, the photos win.
• Buttons are locked separately: visible buttons follow the product photos' spacing, colour, shape and finish, so how many appear is set by the product's own spacing, and a close-up of half a placket never crams in the whole garment's buttons. They are never added, removed, recoloured or reshaped.
• Structural parts appear only where the reference shows them: when the reference is a fabric macro, or a crop that shows no buttons, pocket, piping, collar, cuff or waistband, none of them are added, even though the product has them.
• The reference decides how the shot looks, never what the product is: composition, framing and crop, whether a model appears and their face, body, hands and pose, the scene, background and props, the lighting, the way details are presented (inset panels, call-out boxes, their number, size and placement), and the surface feel of the cloth, plus where the product sits in frame and how large it is.
• On-image text keeps the reference's position, font, size and colour; only words naming the old product's colour or pattern change to fit the new SKU, and every other word stays as written.
• Extra requirements reach the model as their own section: where they conflict with the rules above they win, but the product always stays this SKU and never turns back into the reference's product.
• The product's colour, pattern and structure are never taken from the reference: for a check or a stripe, the width and order of every band, the colours inside it, the direction and the proportions of the repeat all come from that SKU's own product photos. The reference's own pattern is discarded entirely: not a single line of it survives, even when fabric fills the whole frame.
• A fabric detail follows only the reference's way of shooting: the fold, stack, exposed edge or seam, magnification, crop, angle and lighting. The pattern is rebuilt from the product photos and laid into the reference's folds; only the cloth's surface feel, such as brushed nap, softness and sheen, may follow the reference.
• Product photos are fact sources, not layout sources: their framing, background, lighting and the way the product is worn or laid out never carry into the result, and whether a person appears is decided by the reference alone — with a person-less fabric reference, the result has no model even when every product photo is worn by one.