Agent Sprite Forge Automates 2D Game Characters From One Approved Still
Agent Sprite Forge 0.4 turns one approved master still into a full six-action sprite set. Pipeline: master still, image-to-video per action, soft-matte key, loop detection, retime, finish, package. Supports OpenAI, Gemini, xAI, BytePlus, fal.ai APIs with Codex/Grok CLI fallback.

Takeaways
- Agent Sprite Forge 0.4 turns one approved master still into a full six-action sprite set.
- Pipeline: master still, image-to-video per action, soft-matte key, loop detection, retime, finish, package.
- Supports OpenAI, Gemini, xAI, BytePlus, fal.ai APIs with Codex/Grok CLI fallback.
- Exports to Godot 4, Aseprite, Tiled, LDtk, plus WebM alpha and packed-alpha MP4 for web.
- Pixel finish drops palette from 700 to under 20 colors with binary alpha.
- MIT licensed, available on GitHub as a Claude Code, Codex, and Grok plugin.
Producing a 2D game character with generative models often leaves developers slicing image sheets, extracting video frames, correcting color drift, and rebuilding transparency by hand. Agent Sprite Forge automates that pipeline: a developer approves a master still, then the agent generates idle, walk, run, attack, jump, and hurt animations before keying, retiming, color-locking, and packaging them for game engines.
The MIT-licensed project runs as an agent skill for Claude Code, Codex, and Grok, meaning those coding agents can invoke its scripts as part of a prompted workflow. Version 0.4 replaces image-sheet slicing with image-to-video generation followed by deterministic Python processing, which makes keying, alignment, timing, and export repeatable. The repository has attracted 4.5k GitHub stars.
One approved still anchors six actions
The workflow begins with a prompt to create the character’s master image. The agent runs master_still.py generate --takes 3 , presents three candidates, and can edit a near-match before the developer approves it. The selected image is padded to a standard framing, such as a 788-pixel-tall hero on a 1024-by-1024 canvas, then recorded in master.json . Every later generation references that file to preserve identity, proportions, and camera distance.
After approval, a second prompt asks for the sprite set. The sprite_set.py pipeline creates one image-to-video clip per action and processes it through several stages:
- Composition checks: character area, foot visibility, zoom, frame edges, background, and unwanted objects.
- Continuity checks: identity, accidental turns, end pose, and color bleed.
- Motion checks: movement quality, action timing, and loop suitability.
- Corrections: failed checks add targeted instructions to the next generation attempt.
- Finishing: soft-matte keying, frame registration, loop selection, one-shot retiming, palette control, and engine packaging.
The resumable run command skips completed clips when a job restarts, avoiding unnecessary generation calls. A soft matte uses partially transparent edge pixels while removing the background, and registration aligns the character between frames to reduce visible jitter.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

