DeepSeekSalesRevenue growthAI score · 76/100AI-assisted edit · no obvious prompt-injection riskGitHub

Reverse-engineer viral e-commerce videos with DeepSeek vision model to generate Seedance replication prompts

Input: viral e-commerce video → Process: frame extraction + S-A-C-S-C shot analysis + styled generation → Output: copy-ready Seedance prompt pack

Some experienceStepsWorkflowTip
Who
E-commerce content creators, short-video directors, teams needing to replicate or adapt viral product videos at scale
You get
Decompose any product video into time-anchored shot cards, then generate three-style (faithful/organic/paid) Seedance-executable prompt packs with negative prompts and reference image lists
You need
DeepSeek

Constraints

Local Python script workflow. Prerequisites: Python environment, pip install openai, ffmpeg in PATH, DeepSeek API key (~/.deepseek_api_key or DEEPSEEK_API_KEY env var). Model: deepseek-v4-flash-vision (vision understanding). Cost: pay-per API call; 1 analyze + 1 generate per style; 10s video analyzes in ~16s, generates in ~24–69s.

What they did

Author liangdabiao developed a two-stage Python skill based on deepseek-v4-flash-vision: analyze.py extracts frames at 1fps, uses ffmpeg scene detection to split e-commerce videos into shots, feeds first/middle/last keyframes of each shot to the vision model for S-A-C-S-C five-dimension decomposition into shot cards, then aggregates global analysis into storyboard.json; replicate.py reads this cache, combines style switch (faithful/organic/paid) and product info, and calls one inference with director-rule knowledge base to generate the full prompt pack.

How it went

Author report: using a 10.1s jimeng commercial as example, decomposed into 4 shots in ~16s; faithful replication successfully translated shot cards into time-anchored per-shot prompts; organic variant automatically added shot E and product-specific negative prompts after detecting original lacked conversion hook. (author_report, unverified)

Worth knowing

No audio analysis: voiceover/BGM inferred from visuals only; accurate voiceover needs manual supplementation. Videos with intense motion or complex multi-person interactions are deliberately slowed/simplified for stability. OCR unreliable for text <20px. Non-e-commerce videos can be decomposed but conversion-hook fields less relevant. Model ID deepseek-v4-flash-vision is author-specified; normalize to deepseek-v4-flash if needed with gap noted.

Steps

  1. Environment setup and repo clone

    Clone the repository locally, install Python dependencies, and confirm ffmpeg is in your system PATH. Write your DeepSeek API key to the first line of ~/.deepseek_api_key, or set it as the DEEPSEEK_API_KEY environment variable.

  2. Run analyze.py to generate storyboard.json

    From the repo root, run python scripts/analyze.py "your-video.mp4". The script extracts frames at 1fps, uses ffmpeg scene detection to split the video into shots, feeds first/middle/last keyframes of each shot to deepseek-v4-flash-vision for S-A-C-S-C five-dimension decomposition into shot cards, then aggregates global analysis. Output is cached to .cache/<video-sha-first-16-chars>/storyboard.json; the same video is never re-analyzed. A 10-second video takes ~16s.

  3. Run replicate.py to produce prompt pack

    Run python scripts/replicate.py "your-video.mp4" for faithful replication. Add --style organic for organic seeding or --style paid for paid-ad conversion. Use --product "brand + selling point" to override AI guesses; use --shot B to regenerate only a specific shot. Each generation makes 1 inference call (max_tokens 32768), taking ~24–69s. Outputs: <video-name>_<style>_prompts.md and .html (with original keyframe comparison).

  4. Execute generation on Jimeng (Seedance) platform

    Prepare 1–3 product photos per the @reference-image list (natural light, solid background, sharpened cutout); blur/crop/illustrate real faces for compliance. Open Jimeng web → Video Generation → Omni-Reference mode, upload assets, paste the full prompt and English negative prompt. For simple scenes (≤2 shots), use Seedance 2.0 fast to save credits; render at 720P first to verify, then 1080P final. If results are off, return to this skill and change one variable (--product, --style, or --shot) rather than restarting from scratch.

Prompt you can copy

Reference prompt · Editorial suggestion。Confirm the destination, input variables and output format before use.

This block is the prompt. Ads are never placed here.

No prompt needed — just follow the steps.

Source

用 DeepSeek 视觉模型拆解带货爆款视频，生成可直接粘贴即梦（Seedance）的复刻提示词包 · liangdabiao · GitHub · 2026-08-26T00:52:55Z

Comments & feedback

Tried this method? Share results, problems or improvements. Comments appear after moderation. Please avoid sensitive information.

Loading…

    Related