threerocks/sweety-handdraw-video

A production pipeline for narrated hand-drawn explainer videos, not a one-click generator: expect to supply an agent host, an image generator and a local runtime, then iterate before anything is publishable.

The problem

You want a short explainer where an illustrated hand draws and writes while a voice walks the viewer through the story. Normally that means writing the script in one tool, generating illustrations in another, drawing them by hand, recording narration, and cutting it to fit audio in an editor. The hard step is timing: matching how long each stroke and caption stays on screen to the spoken words. Here script, images, voice, timestamps and rendering stay inside one skill directory, so revising a caption or layout reuses the unchanged images and audio instead of restarting the video.

What's different

The interesting mechanism is budget-first storyboarding: the pipeline computes how many seconds and strokes each illustration can afford from the narration, then designs images to that budget rather than squeezing the timeline afterward. Narration is generated per semantic paragraph, each segment passes hiss/noise/clipping gates, and timestamps drive when drawings and captions appear. Drawing mimics a person: ink-outline styles are stroked fully, contour before detail, with pen lifts; marker shading goes in zigzags; Chinese text follows standard stroke order.

That yields a more locatable result than a single generation pass: scripts, assets, audio, timestamps and check records are kept, so subtitle or layout revisions re-render without regenerating images. The trade-off is that this is a workflow, not a button. It depends on an agent to write narration and storyboards and on an image-generation tool for illustrations. Validation is narrow: the full film pipeline is said to be tested on Apple Silicon Mac with Codex; Claude Code has verified installation and local scripts but the whole-video flow is untested there, and Windows is not promised. See the installation instructions and the sample films.

Who should try it

Good fit: someone producing teaching, product-principle or fable-style explainers on a Mac who already has an agent host and an image generator and is willing to inspect each result. The revision model suits creators who expect to tweak wording, subtitles or pacing repeatedly. Poor fit: anyone wanting unattended batch output for arbitrary inputs, since the author says every video needs checking for character, story or voice drift. Windows users should not start here, and those without a local TTS path can use a prepared narration track with timestamps.

Requirements and limitations

Installation is roughly 38 MB and offline-ready for hands, style previews and samples, but setup writes a Python virtual environment outside the skill directory and installs FFmpeg via Homebrew or apt. Voice is separate from voice tone: the local VoxCPM2 adapter needs Apple Silicon and about 3.2 GB of model download; the VoiceStudio local HTTP service needs a configured TTS engine; the optional 火山豆包 / Gemini TTS path was only tested against simulated HTTP, not real calls. Updates are not automatic, so replacing the installed folder requires re-checking the bundled VERSION. Ongoing work means checking audio by ear and declaring AI-generated content on publishing platforms despite embedded metadata. The author's "about 10 minutes" and "under 1% quota" are a specific self-report, not a fixed price.

HotRepo's take

HotRepo first recorded this repo on 2026-10-24 and saw 842 total stars at #14 on the NEW-REPOSITORY board for JavaScript on 2026-10-25 (a language-filtered sample, not an overall Trending rank or a daily-growth figure). With only two recorded dates and no comparable daily-board history, persistence cannot be judged. Before adopting, check that your agent host and image generator pass the doctor checks for execution, image generation, fonts and voice on your machine; the FAQ covers the supported combinations.