dasheng-video-explainer-html

name	dasheng-video-explainer-html
description	Use when turning a Dasheng HTML article into a no-human vertical explainer video with voiceover, storyboard scenes, HTML animation, real charts, music, and final MP4.

Dasheng Video Explainer HTML

Role

Build the无真人出镜科普 production line. The article is the fact source; the storyboard is the first-class artifact; html-video renders the animated scenes.

Required First Artifact

Build the template router, extract the article into a storyboard, then expand it into an HTML Anything video timeline:

python3 scripts/build_html_anything_template_router.py \
  --output configs/video/html_anything_template_router.json

python3 scripts/video_explainer_storyboard.py \
  --html <article.html> \
  --template-router configs/video/html_anything_template_router.json \
  --output <explainer_storyboard.json> \
  --preview-html <storyboard_preview.html>

python3 scripts/build_storyboard_template_review_table.py \
  --storyboard <explainer_storyboard.json> \
  --output <storyboard_template_review.html>

python3 scripts/extract_template_preview_frames.py \
  --template-data <template_showcase_data.json> \
  --video <template_showcase_silent.mp4> \
  --output-dir <template_previews_dir>

python3 scripts/build_html_anything_video_timeline.py \
  --storyboard <explainer_storyboard.json> \
  --article-html <article.html> \
  --template-router configs/video/html_anything_template_router.json \
  --output <html_anything_video_timeline.json>

Optional style asset:

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

Use it to tune template preferences, scene length, transitions, color palette, title/opening/outro rhythm, Lottie/GSAP density, and evidence-scene spacing. Do not train on sample videos inside this production step; use dasheng-video-style-trainer first.

Director Mechanism

Read docs/technical/video-editing-driving-mechanism.md, docs/technical/video-script-template-routing-guide.md, and configs/video/video_editing_driver_rules.json before rendering.

No-human explainer is not PPT pagination. Drive scene choice with:

voiceover beat -> evidence_need -> cognitive_load -> template -> intra-scene motion -> transition -> music/SFX

Default state machine:

hook_card -> question_setup -> chapter_card -> evidence_scene -> logic_animation -> cinematic_bridge -> evidence_scene -> recap_card -> outro

Production Rules

• Do not create a second fact chain. Reuse the article's real tables, charts, images, and sourced claims.
• Build a source-asset inventory before writing the storyboard. Original article charts are mandatory evidence assets: none may disappear between article intake, storyboard, scene pack, and final render without an explicit approved exclusion reason.
• Convert each retained article chart into a dynamic chart sequence: source evidence hold -> one-time data redraw -> completed-chart reading hold. Never use static zoom/pan, decorative replacement data, or repeated entrance/exit loops.
• Scene duration must follow cognitive load. Default guidance: 2-4 seconds for a simple transition, 4-7 seconds for a single claim, and 8-15 seconds for a dense chart/table. Long evidence scenes should gain internal micro-shots rather than being fragmented into unreadable short scenes.
• Adjacent sentences that share one visual argument should remain in the same scene. Repeatedly cutting back to the same card or making one layout enter/exit multiple times is a director failure.
• Show the overall outline once. Chapter scenes may use a persistent progress marker, but must not alternate between the full outline and enlarged sub-outline cards.
• Learned video style controls presentation only. It must not invent facts, replace article charts, or copy sample-video scripts/frames.
• Generate narration with MiniMax CLI (mmx). For review/final videos, prefer per-scene TTS or provider timestamps so each scene duration comes from real audio duration. A single continuous TTS file with text-length timing is only acceptable for rough preview.
• Default no-human narration is tianxin_xiaoling at 1.2x speed. If narration feels slow, compress the visual timeline by the same ratio as the voice; do not speed up voice alone and leave visuals drifting.
• Use MiniMax CLI for production narration, background music, AI image assets, and generated口播音频. Do not call MiniMax APIs directly from ad-hoc scripts unless the CLI cannot express the operation.
• Use external html-video as the default renderer via dasheng-html-video-bridge; install on demand with scripts/ensure_video_external_deps.py.
• Use external html-anything only as visual/template reference via dasheng-html-anything-bridge; install on demand with scripts/ensure_video_external_deps.py.
• No generic framework diagrams when article data can support a concrete chart/table.
• Charts, tables, and line graphs must be backed by article data, source images, or verified data pulls. Never draw decorative or fake lines for a data scene.
• Template diversity must be real at the renderer/component level, not just a template title label.
• Reserve safe zones before rendering: keep core charts/tables away from top chrome and bottom captions; no card/table overlap, no element collision, no developer-facing labels in final MP4.
• Each finance/data scene must answer “what does this prove?” on screen. Avoid visually busy but content-light template demos.
• Final captions must cover the full voiceover, not just scene summaries. Captions should be generated as timed JSON/SRT cues, displayed in sync with the active spoken sentence, and included as delivery artifacts.
• Caption display text must normalize years, percentages, quantities, and counts to Arabic numerals where readability is better, e.g. 2022-2025, 50%, 3个月, 5句话.
• Subtitle timing must be audio-driven. Use per-scene audio durations, provider alignment metadata, or ASR/forced-alignment backfill. If subtitles are generated only by proportional text length, label the render as preview and do not deliver it as final.
• When only a continuous TTS file exists, run scripts/align_video_subtitles_to_asr.py --project-dir <remotion_project> --asr-json <whisper_json> --speed 1.2 --write, then re-render. This keeps script text but uses ASR timestamps.
• No CDN-dependent final assets; preview HTML may be simple, final MP4 must be rendered locally.
• Prefer HyperFrames as the scene composition model.
• Use GSAP-style timelines for entrance, exit, chart reveal, path draw, table scan, and title kinetics.
• Lottie is allowed only as decorative/auxiliary motion. It must never replace real article charts, tables, screenshots, or evidence.
• When a scene needs a reusable vector motion accent, route it through dasheng-html-video-bridge and optional external text-to-lottie: warning icon, document scan, data flow, market ticker, path trace, lower-third, chapter symbol, or outro mark. Generated Lottie JSON must be verified in the Skia Skottie player and recorded under the task output folder before final render.
• Final/review videos must be rendered by live HTML animation recording. Static screenshot, PNG stitching, or FFmpeg zoompan is forbidden for production because it destroys GSAP/Lottie/template motion.
• Do not render from raw storyboard directly unless debugging. Final video planning must route content parts to HTML Anything templates first.
• Before TTS, material generation, or video render, produce storyboard_template_review.html and ask for user approval. This page must show one row per scene with time, voiceover, core meaning, evidence refs, template id, template screenshot/placeholder, motion plan, risk notes, and review decision controls.
• Require the user/exported storyboard_review_decision.json before production. Validate it with scripts/validate_storyboard_review_gate.py --storyboard <storyboard.json> --decision <storyboard_review_decision.json> --output <storyboard_review_gate.json>.
• Do not proceed if the gate report has status != approved or render_allowed != true.
• If a template has no screenshot, show a visible “暂无模板截图” placeholder and the required preview path. Do not fake a screenshot or hide the missing preview.
• Prefer real template screenshots from scripts/extract_template_preview_frames.py, template preview.png, or a renderer still. Do not substitute unrelated scene frames as template screenshots unless the table explicitly labels them as scene previews.
• macOS say is only a smoke-test fallback; it is not acceptable for final voiceover unless explicitly requested.
• Evidence scenes must appear every 20-35 seconds. Chapter or structure reset must appear every 45-90 seconds.
• A scene longer than 8 seconds needs explicit intra-scene motion: data reveal, document zoom, path highlight, focus shift, or exit motion.

MiniMax CLI Defaults

Check authentication and quota before production rendering:

mmx auth status --no-color
mmx quota --no-color

Default render command shape:

python3 scripts/render_html_anything_scene_pack_animated.py \
  --manifest <scene_pack_manifest.json> \
  --output-dir <render_output_dir> \
  --with-voice \
  --voice-provider mmx \
  --voice "tianxin_xiaoling" \
  --mmx-model speech-2.8-hd \
  --mmx-speed 1.2

Default narration command shape used by the renderer:

mmx speech synthesize \
  --text-file <full_voiceover_script.txt> \
  --out <voiceover_single.wav> \
  --model speech-2.8-hd \
  --voice "tianxin_xiaoling" \
  --speed 1.2 \
  --format wav \
  --sample-rate 44100 \
  --channels 1 \
  --language Chinese

Default background music command shape:

mmx music generate \
  --prompt "light technology explainer, data reveal, restrained financial documentary, no vocals" \
  --instrumental \
  --out <bgm.mp3>

Default AI image command shape:

mmx image generate \
  --prompt "<article-specific visual prompt>" \
  --aspect-ratio 9:16 \
  --out <image.jpg>

Style Targets

• Vertical finance/documentary style for mobile publish; horizontal 16:9 is acceptable for Bilibili, template review, and sample evaluation unless the channel requires vertical.
• Default voice: MiniMax tianxin_xiaoling, speed 1.2, warm investor-chat delivery. Keep rhetorical pauses in the script but avoid slow TTS pacing.
• Default BGM: light technology explainer / data reveal. Keep BGM low under the voice and use chapter risers sparingly.
• Average scene: 5-7s; median: 4-5s.
• Evidence screen every 20-35s.
• Chapter card every 45-90s.
• Prefer document zoom, data reveal, chart animation, terminal/Bloomberg-like information rhythm.
• Add Lottie-style accent motion for warning, market ticker, data flow, document scan, and outro only when it supports the spoken beat.
• For AI-generated Lottie assets, prefer prompts grounded in real article variables: exact label text, numeric value, direction of movement, target duration/FPS, transparent background, and where it will sit in the scene. Avoid vague prompts like “make it professional”.

Review Feedback Guardrails

• If a reviewer flags “line charts are sloppy,” first check whether numeric columns were parsed incorrectly as dates; fix data extraction before visual styling.
• If a reviewer flags “穿模/遮盖,” reduce component density and regenerate midpoint contact sheets before final render.
• If narration is too slow, target 1.2x voice speed and rebuild scene duration allocation from the actual audio duration.
• If a template review video looks repetitive, reject title-only template switching and require one unique renderer behavior per template.
• If source charts are missing, reject the render even when technical QC passes. Compare the source-asset inventory with the final scene manifest and require 100% accounted-for coverage.
• If viewers cannot finish reading a chart before the next cut, extend its completed-state hold; do not solve the problem by replaying the chart entrance.
• If captions are missing content or lag voice, reject the render: rebuild timed subtitle cues from the full voiceover plus real audio timing, then re-render a contact-sheet/video sample.

Output Contract

• explainer_storyboard.json
• html_anything_video_timeline.json
• storyboard_preview.html
• storyboard_template_review.html
• storyboard_review_decision.json
• storyboard_review_gate.json
• voiceover.wav or provider-specific audio file
• captions_full.json
• captions_full.srt
• final_explainer_vertical.mp4
• video_qc_report.json
• qa_contact_sheet.jpg