akachandan1/video-studio

Art-style animation → HyperFrames

Engines

• huashu-art-motion (~/.claude/skills/huashu-art-motion, installed by setup.sh): deterministic canvas engine (window.renderFrame(t)), painting libs PAINT (strokes, halftone, pixelate, dabs, mosaic, facets, inkWash, brush, cut → Path2D), TOON (flat shading), CAM (pushTo dives). Read its SKILL.md for the API.
• claude-animation plugin (hand-drawn 2D, textured-editorial / watercolour looks) — alternative look.
• HyperFrames does assembly, titles, sound, captions and the final render.

Workflow that works

1. Custom scenes, not presets. huashu's render.py --spec clips draw a centre node + satellites and read as a mind map. For anything showcase-grade write scenes: copy scripts/engine into videos/<name>/engine, create demos/<film>/eras.js + <film>.js registering SCENES['id'] = (g, t, W, H) => {...}.
2. Kurzgesagt look: one continuous world, flat shapes with TOON shading, rim light, a slow CAM.pushTo dive into each layer ("what's inside"), Poppins/rounded type, scale bars and labels that pop on the beat.
3. Painter styles: compose the subject once, then render it through each style's brush (Van Gogh = directional strokes along a flow field; Lichtenstein = halftone + black outline; Kusama = dots; cubism = facets ~nx26/ny16 with small shift — too many facets = shattered mess; ink = inkWash; 8-bit = pixelate + palette).
4. Render (30 fps, the engine defaults to 60): uv run --with playwright python render.py --film demos/<film> --solo <id> --no-counter --fps 30 --from 0 --to <s> --out clip.mp4 (--width 1080 --height 1920 for vertical).
5. Assemble in HyperFrames: clips as <video class="clip"> with titles/labels as HTML on top, score from the score skill with a hit per style change, render, loudnorm −14.

Look at a frame of every style before rendering the full set — alpha/wash settings that look fine in code often come out muddy.