dasheng-video-style-trainer

name	dasheng-video-style-trainer
description	Use when learning reusable video style DNA from many sample videos, training a blogger/channel editing style, or applying a learned style profile to Dasheng talking-head and no-human explainer video workflows.

Dasheng Video Style Trainer

Role

Build an optional video-learning asset before production. It analyzes user-provided reference videos and produces a reusable style_profile.json for later video generation.

This is not a mainline gate. The canonical chain remains:

intake -> brief -> draft -> transwrite -> publish -> postmortem

Video training is an optional asset generator like article paradigm learning. Use it when the user provides many sample videos, names a target creator style, or asks future videos to follow a learned editing/visual rhythm.

Storage Rule

All videos, upload caches, training reports, rendered previews, audio, subtitles, and other generated media must stay outside the repo.

Default output root:

~/Desktop/自媒体创作/00_范式学习/视频训练/<style_id>/

Never write source videos or training outputs to:

• skills/
• project root
• openclaw-skill-exports/
• any skill package directory

Source videos may stay wherever the user provided them. The training manifest records their paths; it does not copy them into the repo.

Train Style DNA

Default command:

python3 scripts/run_mainline_stage.py video-train \
  --video-dir <sample_video_dir> \
  --style-id <style_id> \
  --blogger-name "<name>" \
  --blogger-platform "<platform>"

Direct command:

python3 scripts/learn_blogger_video_style_local.py \
  --video-dir <sample_video_dir> \
  --style-id <style_id> \
  --blogger-name "<name>" \
  --blogger-platform "<platform>" \
  --output-dir ~/Desktop/自媒体创作/00_范式学习/视频训练

Requirements:

• Prefer local claude-real-video for first-pass video reading: scene-aware keyframes, contact sheets, manifest, and optional transcript.
• Do not upload reference videos to model providers in this stage.
• Use ffprobe and ffmpeg for metadata and oversized-video compression.
• Keep compressed upload copies under <style_id>/_upload_cache/.
• Use --skip-aggregate only for debugging single-video analysis.

Local video reading command:

python3 scripts/read_video_with_crv.py <video_or_url> \
  --output-dir ~/Desktop/自媒体创作/00_范式学习/视频训练/<style_id>/crv/<video_stem> \
  --why "提取剪辑节奏、镜头构图、证据画面、转场和风格 DNA" \
  --report

Use --transcribe only when local Whisper is installed and transcription is needed. Do not store crv-out in the repo.

Output Contract

Training must emit:

• training_manifest.json
• optional crv/<video_stem>/dasheng_video_reading_manifest.json
• per_video/<video_stem>/analysis.json
• style_profile.json unless --skip-aggregate is used
• style_profile.md unless --skip-aggregate is used
• _upload_cache/*.upload.mp4 only when a source video exceeds upload limits

training_manifest.json must record:

• stage: video-style-training
• style_id
• sample video paths
• output paths
• model name
• status
• whether source videos were copied
• where compressed upload cache lives

Apply Style

Before selecting a profile, read configs/video/reference_video_dna_registry.json. Use style_profile.curated.json when the registry points to one; the automatically aggregated style_profile.json is only the statistical fallback.

The registry distinguishes:

• fixed user-approved standard samples;
• rolling "latest creator video" samples with an explicit checked_at date;
• classic benchmarks that must not be silently replaced by a creator's latest format.

When refreshing a rolling sample, create a new dated training directory and registry entry. Do not overwrite an approved historical profile in place.

Use a learned style_profile.json to drive HTML Anything and html-video timelines:

python3 scripts/apply_blogger_style_to_timeline.py \
  --style-profile ~/Desktop/自媒体创作/00_范式学习/视频训练/<style_id>/style_profile.json \
  --storyboard ~/Desktop/自媒体创作/06_转写生产/<run_id>/<topic>/video_storyboard.json \
  --article-html ~/Desktop/自媒体创作/05_初稿生成/<run_id>/<topic>/03_HTML草稿_<topic>.html \
  --output-timeline ~/Desktop/自媒体创作/06_转写生产/<run_id>/<topic>/html_anything_video_timeline.style.json

End-to-end preview:

python3 scripts/run_blogger_style_video_pipeline.py \
  --style-profile ~/Desktop/自媒体创作/00_范式学习/视频训练/<style_id>/style_profile.json \
  --storyboard ~/Desktop/自媒体创作/06_转写生产/<run_id>/<topic>/video_storyboard.json \
  --article-html ~/Desktop/自媒体创作/05_初稿生成/<run_id>/<topic>/03_HTML草稿_<topic>.html \
  --output-dir ~/Desktop/自媒体创作/06_转写生产/<run_id>/<topic>/blogger_style_video

Consumption Rules

• Talking-head videos use learned pacing, evidence interval, subtitle density, return-to-speaker rhythm, transition signature, and overlay placement.
• No-human explainer videos use learned template preferences, motion style, color palette, title/opening/outro patterns, and B-roll/data-scene density.
• Draft remains the fact source. Style training must not invent market data, charts, or claims.
• Style DNA controls form; article paradigm controls structure; account DNA controls voice.

Quality Gate

Before using a style profile downstream, check:

• At least 3 representative sample videos unless the user explicitly requests a one-shot experiment.
• style_profile.json has visual_dna, editing_dna, and template_preferences.
• talking_head_policy.blank_placeholder is true for reference creators; user footage replaces the placeholder later.
• No generated media exists in repo root or any skill folder.
• The style profile describes reusable patterns rather than copying exact copyrighted frames, scripts, or slogans.
• The selected profile's scope matches the production lane. A sponsored interview must not replace a classic finance-explainer benchmark.
• A curated profile records both strengths_to_reuse and limitations_not_to_copy.