Bilibili AI Video Prompt Analyzer#

Extract/reverse-engineer production-ready prompts from Bilibili AI-generated videos for imitation. Outputs detailed prompts in ComfyUI/Seedance format that can be used directly with minimal edits.

Workflow#

1. Download video: Get video from Bilibili URL using yt-dlp

2. Extract subtitles: Extract soft or hard subtitles for voiceover reference/rewrite

3. Extract frames: Extract frames at configurable FPS (default: 1 frame/second)

4. Visual analysis: Analyze frame-by-frame to identify style, subjects, colors, composition, motion changes

5. Structure output (ready for direct use):

   • Technical metadata (aspect ratio, duration, fps, model recommendation)
   • 2-3 keyframe prompts (most distinct compositions, ComfyUI comma format)
   • Scene sequence by story beat (timeline + prompt evolution)
   • Full video master prompt (overall style + theme)
   • Extracted subtitle text (for voiceover reference)

When to Use#

✅ Use this skill when:

• You have a Bilibili link to an AI-generated video
• You want to imitate the style and generate similar videos
• You need production-ready prompts that work directly in Seedance/ComfyUI/Kling
• You want to extract subtitles to reuse or rewrite the voiceover
• You want keyframe + scene breakdown for multi-step generation

❌ Don't use when:

• The prompt is already in the video description (use agent-reach to extract directly)
• It's a human-shot video not generated by AI

Sampling Strategy#

FPSUse CaseRecommendation0.2Long video (>5min), big scene changesOnly key frames1Short video (<1min)Default, one frame per second2-5Detailed motion analysisSlow morphing animations

Best practice: 1fps for short videos (<1min) gives good granularity without redundancy. Most AI video evolution changes gradually, 1s interval captures the arc well.

Requirements#

• ffmpeg installed (for frame extraction and subtitle extraction)
• yt-dlp installed (for video download)
• Multimodal LLM support (for visual frame analysis)

Scripts#

• scripts/download-bilibili.sh <url> <output-dir> - Download video from Bilibili URL
• scripts/extract-frames.sh <video-file> <output-dir> [fps] - Extract frames from video
• scripts/extract-subtitles.sh <video-file> <output-path> - Extract subtitles to text file (supports soft and hardsubs)

Output Format (Production-Ready)#

1. Technical Metadata#

text

Aspect Ratio: 9:16 (1080x1440)
Duration: 35 seconds
Frame Rate: 30 fps
Recommended Model: Seedance 2.0 / Kling AI / Runway Gen-3
Motion Type: Slow fluid morphing, seamless transformation

2. Keyframe Prompts (ComfyUI Format, comma-separated)#

One detailed prompt for each most distinct keyframe:

text

masterpiece, best quality, 8k, ultra detailed, tiny glowing point of light in deep indigo mist, volumetric godray, soft atmospheric haze, vertical composition 9:16, surreal abstract, mindscape, glowing particles, deep blue purple color palette, ethereal, dreamlike, cinematic lighting

3. Scene Sequence Breakdown (by timeline beat)#

TimeVisual DescriptionPrompt Addition0s-8sOrigin in voidglowing point, mist, void8s-18sNetwork growthfibers spreading, interconnected thought network

4. Full Video Master Prompt#

Complete cohesive prompt capturing overall style, theme, pacing that can be used directly.

5. Extracted Subtitles#

Clean text extracted for voiceover reference or rewrite.

Prompt Quality Guidelines#

Output prompts follow Seedance/ComfyUI best practices:

• Start with quality tokens: masterpiece, best quality, 8k, ultra detailed
• Describe composition before style
• Specify exact colors and lighting
• Include aspect ratio and motion characteristics
• Keep comma-separated format for direct copy-paste