Video2DSprite takes an approved master still and makes a clip for each action (idle, walk, run, attack, jump, hurt, cast and more) with an image-to-video model. It then gates each take with numeric checks, keys it with a soft matte, registers it to the master, picks loops or retimes one-shots to game length, finishes it and packages fixed-canvas frames.
Packages include alpha WebM, a packed MP4 for iPhone, a PNG fallback, engine metadata and a JS runtime. It also processes a clip you supply.
What You Get
- A set plan with a prompt, canvas and timing for every action
- Keyed, registered frames per action, with loops chosen by measurement and one-shots retimed
- A review step with sheets, final images and a lineup to inspect before you accept each action
- Packages: PNG frames, alpha WebM, packed-alpha MP4, engine metadata and a JS runtime
Common Use Cases
- Animate a hero with idle, walk, run, attack, jump and hurt from one still
- Convert a supplied video clip into transparent registered frames
- Retime an attack to a short hit window for game feel
- Package animation for Godot and the web
What Makes It Unique
The numeric gates cover area, feet, identity, zoom, turning around, edges, background, extra objects, motion, end pose, color bleed and timing, and a failed take is retried with a targeted fix clause. The skill is candid that video is not promised to be smaller, seamless or cheaper than sheets, and it asks you to look at every result because numbers cannot see identity or a changed weapon.
Under The Hood
The agent runs Python scripts from the skill folder through the shell and uses ffmpeg to process video. Clips come from the shared media route command, which sends the still and prompt over the network to a video provider you configured by API key (xAI, BytePlus, fal.ai) or to your Grok command line tool. Scripts write many new files and folders in your project and can resume a run. Read the scripts before running them on a sensitive machine; the automated scan findings are shown on the page.