dsh-animation-studio

Producing a 30-second explainer on gradient descent traditionally means dragging tracks in After Effects or hand-coding hundreds of lines of Remotion — a high bar for most educators and content creators. dsh-animation-studio turns that into a single sentence in a DSH chat: state the topic, length, and pacing, and the AI walks through storyboard, scene-by-scene composition, frame preview, timeline fine-tuning, and full MP4 render; you can undo, swap easings, or nudge keyframes at any point without touching a line of code.

The key design decision: the model only reads and writes a single AnimationSpec JSON timeline and never generates animation code. All timing uses scene-relative absolute milliseconds, animatable attributes live on track keyframes, and easings cover linear, spring, bounce, elastic, and more. Every edit is a JSON Patch with a reversible inverse, so undo costs zero re-inference. Ten anim_* tools split the pipeline into self-diagnose, plan, draft, asset-import, preview, and render — each response carries a next-step hint. The render layer silently handles Motion Canvas headless capture, ffmpeg muxing, and CJK fonts out of the box; scene-level caching lets unchanged scenes skip re-rendering. Narration cues produce auto-sized subtitles, and optional TTS (edge-tts or piper) mixes voice into the final clip. The renderer itself is a pluggable seam, so swapping in Remotion or Manim means writing one Provider while tools and event flow stay untouched.

Who it fits: teachers who need a 30-second to 2-minute explainer quickly; content creators who want to iterate on pacing and easing inside a conversation instead of opening a timeline editor; and DSH power users who expect a rich-card workbench in the Web UI — every tool call renders as a visual card, finished videos play inline with scrub and download, and background jobs show a live progress bar.