Turn static product images into 3–6 second "moving shots": each image gets one image-to-video prompt + one local fallback solution (ffmpeg Ken Burns), ensuring you can produce a video even without a video model.
Part of ecommerce-video-skills · Where do the images come from? Use the sister repository ecommerce-image-skills to first produce white-background images/scene images/selling-point images.
Inputs
| Required | Description |
|---|---|
| 1–6 product images | White-background main images, scene images, detail images are all fine; the clearer the better (≥1000px) |
| Placement platform | Douyin / TikTok Shop / WeChat Channels / Taobao main image video / Amazon (determines aspect ratio 9:16, 1:1, 3:4, 16:9) |
| Optional | Selling points (up to 3), target duration, style (fresh / premium / energetic), available video models |
Steps
- Product consistency lock (mandatory): Look at the image and write down shape, color, Logo/text, material, accessories—then carry them verbatim in every prompt,forbid the model from changing the product。
PRODUCT LOCK: tall matte sage-green bottle, natural bamboo lid with black loop,
small white "AURA" wordmark on upper body. Keep shape, colour, logo and proportions identical in every frame.
-
Choose one shot role for each image(one image does only one thing):
Image type Recommended shot Action White-background main image Slow push-in / 360° orbit (when the model supports it) Product stays still, light and shadow sweep across Scene image Slight lateral movement / hand enters frame A person picks it up, pours water, puts it in a bag Detail image Macro push-in Light spots sweep across the material texture Selling point/infographic Static or extremely slow push Do not let the text deform (keep the text area as still as possible) -
Write the prompt(template below): subject → action → camera movement → lighting/atmosphere → duration/aspect ratio → constraints. The smaller the action, the more stable the product.
-
Give every shot a local fallback: when there is no video model / the model deforms the product, use
ffmpeg-auto-assemble'smotionfield (zoom_in/pan_left/punch…) to directly make the still image move. -
Deliverable: shot table (number, image, prompt, fallback motion, duration) + shots segment that can be pasted directly into
storyboard.json.
Prompt templates
General image-to-video (Chinese)
[参考图] 作为首帧。{PRODUCT LOCK}
镜头：{缓慢推近 / 从左向右平移 / 低角度环绕 30°}，运动平稳，无抖动。
画面动作：{冰块落入杯中溅起细小水花 / 手从画面右侧入画拿起杯子 / 柔和光斑从左扫到右}。
光线：{明亮自然光 / 柔和棚拍光，浅景深}。
时长约 {4} 秒，{9:16} 竖屏。
要求：商品形状、颜色、Logo 与参考图完全一致；不新增文字；不出现其他品牌；手部自然、五指正常。
English (most models parse English best)
Use the reference image as the first frame. {PRODUCT LOCK}
Camera: {slow push-in | gentle left-to-right truck | 30° low-angle orbit}, smooth, no shake.
Action: {ice cubes drop into the bottle with a small splash | a hand enters from the right and lifts the bottle | a soft light sweep across the surface}.
Lighting: {bright natural daylight | soft studio light, shallow depth of field}.
About {4} seconds, {9:16} vertical.
Keep the product's shape, colour and logo identical to the reference. No new text, no other brands, natural hands.
First and last frame mode(when the model supports first frame + last frame): first frame = white-background main image, last frame = scene image, and the prompt only says "transition from studio shot to outdoor scene, product remains unchanged, camera slowly pulls back".
Negative constraints (when the model supports negative prompt)：deformed product, melting, extra logos, changed colour, warped text, extra fingers, flicker, watermark
Local fallback (no model)
{"image": "assets/white-bg.webp", "motion": "zoom_in", "line": "换这只保温杯试试。"}
python scripts/assemble.py storyboard.json -o out/main-video.mp4 --size 1080x1080 # 本目录已附带 assemble.py
Checklist
- Every prompt carries PRODUCT LOCK
- One shot has only one main action; the fewer actions the product itself has, the better
- Selling-point images/images with text use only static or extremely slow camera movement
- Aspect ratio matches the platform (main image videos commonly use 1:1 / 3:4, information feeds 9:16)
- Do not write parameters the model does not support; duration and resolution are subject to the documentation of the model used
- Every shot has a local motion fallback
Common Questions and Fixes Failure fixes
| Issue | Fix |
|---|---|
| Product distortion / Logo blurring | Reduce motion, change to "camera moves, product stays still"; or switch to local Ken Burns |
| Hand deformities | Remove hand movements, or have the hands appear for only half a second at the edge of the frame |
| Screen flickering | Shorten the duration, reduce the motion amplitude, turn off exaggerated lighting effects |
| Garbled text | Do not let the model generate text; leave subtitles and decorative text to subtitles-and-text-effects post-production |
| Color drift | Write the color value or color name in the prompt, and require "match the reference colour exactly" |