concept-to-video（概念到视频）

Concept to Video

Creates animated explainer videos from concepts using Manim (Python) as a programmatic animation engine.

Reference Files

Why Manim as the engine

Manim is the "SVG of video" — you write Python code that describes animations declaratively, and it renders to MP4/GIF at any resolution. The Python scene file IS the editable intermediate: the user can see the code, request changes ("make the arrows red", "add a third step", "slow down the transition"), and only do a final high-quality render once satisfied. This makes the workflow iterative and controllable, exactly like concept-to-image uses HTML as an intermediate.

Workflow

Concept → Manim scene (.py) → Preview (low-quality) → Iterate → Final render (MP4/GIF)

1. Interpret the user's concept — determine the best animation approach
2. Design a self-contained Manim scene file — one file, one Scene class
3. Preview by rendering at low quality (-ql) for fast iteration
4. Iterate on the scene based on user feedback
5. Export final video at high quality using scripts/render_video.py

Step 0: Ensure dependencies

Before writing any scene, ensure Manim is installed:

# System deps (usually pre-installed)
apt-get install -y libpango1.0-dev libcairo2-dev ffmpeg 2>/dev/null

# Python package
pip install manim --break-system-packages -q

Verify with: python3 -c "import manim; print(manim.__version__)"

Step 1: Interpret the concept

Determine the best animation pattern, then read the matching rule file before writing any code.

Step 2: Design the Manim scene

Template-first vs from-scratch

Check whether a parametric template covers the concept before writing a scene from scratch:

When using a template: copy it to the working directory, edit the config constants at the top, do not restructure the class.

Core rules:

• Single file, single Scene class: Everything in one .py file with one class XxxScene(Scene).
• Self-contained: No external assets unless absolutely necessary. Use Manim primitives for everything.
• Readable code: The scene file IS the user's artifact. Use clear variable names, comments for each animation beat.
• Color with intention: Use Manim's color constants (BLUE, RED, GREEN, YELLOW, etc.) or hex colors. Max 4-5 colors. Every color should encode meaning.
• Pacing: Include self.wait() calls between logical sections. 0.5s for breathing room, 1-2s for major transitions.
• Text legibility: Use font_size=36 minimum for body text, font_size=48+ for titles. Test at target resolution.
• Scene dimensions: Default Manim canvas is 14.2 × 8 units (16:9). Keep content within ±6 horizontal, ±3.5 vertical.

Animation best practices

# DO: Use animation groups for simultaneous effects
self.play(FadeIn(box), Write(label), run_time=1)

# DO: Use .animate syntax for property changes
self.play(box.animate.shift(RIGHT * 2).set_color(GREEN))

# DO: Stagger related elements
self.play(LaggedStart(*[FadeIn(item) for item in items], lag_ratio=0.2))

# DON'T: Add/remove without animation (jarring)
self.add(box)  # Only for setup before first frame

# DON'T: Make animations too fast
self.play(Transform(a, b), run_time=0.3)  # Too fast to read

Structure template

from manim import *

class ConceptScene(Scene):
    def construct(self):
        # === Section 1: Title / Setup ===
        title = Text("Concept Name", font_size=56, weight=BOLD)
        self.play(Write(title))
        self.wait(1)
        self.play(FadeOut(title))

        # === Section 2: Core animation ===
        # ... main content here ...

        # === Section 3: Summary / Conclusion ===
        # ... wrap-up animation ...
        self.wait(2)

Step 3: Preview render

Use low quality for fast iteration:

python3 scripts/render_video.py scene.py ConceptScene --quality low --format mp4

This renders at 480p/15fps — fast enough for previewing timing and layout. Present the video to the user.

Step 4: Iterate

Common refinement requests and how to handle them:

Step 5: Final export

Once the user is satisfied:

python3 scripts/render_video.py scene.py ConceptScene --quality high --format mp4

Quality presets

Format options

Delivering the output

Present both:

1. The .py scene file (for future editing)
2. The rendered video file (final output)

Copy the final video to /mnt/user-data/outputs/ and present it.

Step 5.5: Optional audio overlay

If the user provides audio (music or voiceover), or requests it:

# Background music at 25% volume with fade-in/out
python3 scripts/add_audio.py final.mp4 music.mp3 \
    --output final_with_audio.mp4 \
    --volume 0.25 --fade-in 2 --fade-out 3 --trim-to-video

# Voiceover at full volume, trimmed to video length
python3 scripts/add_audio.py final.mp4 voiceover.mp3 \
    --output final_narrated.mp4 --trim-to-video

For voiceover scripting before recording, read references/rules/voiceover-scaffold.md. For subtitles/captions, read references/rules/subtitles.md. For advanced multi-track mixing, read references/rules/audio-overlay.md.

Error Handling

LaTeX fallback

If LaTeX is not available, avoid MathTex and Tex. Use Text with Unicode math symbols instead:

# Instead of: MathTex(r"\frac{1}{n} \sum_{i=1}^{n} x_i")
# Use:        Text("(1/n) Σ xᵢ", font_size=36)

Agentic Mode (Opt-In)

Single-shot mode (default) is fast and cheap — the coder writes scene.py directly from a concept. Use agentic mode for production-quality renders where layout correctness and asset resolution matter enough to justify additional LLM and VLM calls.

Pipeline

concept
  └─► plan_storyboard.py ──► storyboard.json
            │
            ▼
      fetch_assets.py (optional)
            │
            ▼
      coder writes scene.py
            │
            ▼
      render_video.py --max-fix-attempts N
            │  ▲
            │  └─ LLM fixup loop (on failure, up to N retries)
            ▼
      critic_pass.py --critic
            │  ▲
            │  └─ VLM layout patch (1 call with M image blocks)
            ▼
       final MP4

Flag Reference

Cost Tradeoffs

The fixup loop adds one LLM call per failed render attempt — with --max-fix-attempts 3 you may pay up to 3 extra calls before the loop exhausts or succeeds. The critic pass adds one VLM call containing N PNG image blocks (default 5, max 10); each frame adds roughly 1 token per 800 bytes of base64-encoded PNG, so complex scenes at high resolution are materially more expensive. Setting --critic-budget to a conservative token ceiling (e.g. 20000) causes BudgetExceededError before the API call is made, so you never pay for an accidentally oversized request — the error is loud and non-recoverable by design.

Invocation Example

# 1. Plan
python3 scripts/plan_storyboard.py "explain transformer self-attention" \
    --output storyboard.json

# 2. (Optional) Fetch assets
python3 scripts/fetch_assets.py storyboard.json \
    --adapter local --asset-dir ./assets --output resolved.json

# 3. Coder writes scene.py (Claude writes this from storyboard.json)

# 4. Render with auto-fix
python3 scripts/render_video.py scene.py AttentionScene \
    --quality high --format mp4 --max-fix-attempts 3 \
    --output final.mp4

# 5. Critic pass
python3 scripts/critic_pass.py scene.py final.mp4 \
    --critic --critic-budget 40000 --frames 5

Agentic pipeline design (storyboard planner, auto-fix loop, VLM critic) is adapted from Code2Video (arXiv 2510.01174, MIT). Vendored prompt templates live in references/code2video/ alongside the upstream LICENSE. Full vendoring record, pinned commit, and re-sync policy are tracked in root ATTRIBUTIONS.md.

Limitations

• Manim + ffmpeg required — cannot render without these dependencies.
• Audio is post-render only — Manim renders silent MP4s. Use scripts/add_audio.py to overlay audio after export.
• LaTeX optional — MathTex requires a LaTeX installation. Fall back to Text with Unicode for math.
• Render time scales with complexity — a 30-second 1080p scene with many objects can take 1-2 minutes to render.
• 3D scenes require OpenGL — ThreeDScene may not work in headless containers. Stick to 2D Scene class.
• No interactivity — output is a static video file, not an interactive widget.
• GIF output is silent — audio overlay only works with MP4/WEBM output formats.

Design anti-patterns to avoid

• Walls of text on screen — keep to 3-5 words per label, max 2 lines
• Everything appearing at once — use staged animations with LaggedStart
• Uniform timing — vary run_time to create rhythm (fast for simple, slow for important)
• No visual hierarchy — use size, color, and position to guide attention
• Rainbow colors — 3-4 intentional colors max
• Ignoring the grid — align objects to consistent positions using arrange/align