htdt/godogen/asset-gen

Asset Generator

Assets come from godogen_assets: local generators on this machine's GPU, one command per task, free. Its checkout here is ${GODOGEN_ASSETS}, and its docs are the reference; this page is the TL;DR plus what the game needs on top. Read ${GODOGEN_ASSETS}/README.md once (conventions, a character start to finish, checking results), and a part's README up to How it works before first using its tool.

A character: qwen-image rgba (T-pose) → gen3d → mia-rig --fingers --anim none → add-moves, with gen-moves for custom moves and lipsync for speech. Props stop after gen3d; one a character holds goes on through grip-prop and add-moves --hold, never by an offset tuned in the engine.

• Always pass `-o` into the project; without it, results land in the godogen_assets checkout.
• One GPU job at a time, minutes per call. Run GPU tools as one sequential batch with a long timeout or in the background, never as parallel calls.
• Game budget. gen3d defaults to ~500k triangles and 2048² textures; pass --faces / --tex (30000 / 1024 for a character).
• Style families. Generate one hero asset, derive the rest with qwen-image edit -i hero.png, prompting only what changes.
• Small sprites. A 1024px image downscaled to 64px looks muddy: design display sizes ≥128px, prompt bold flat forms, or generate a kit (several objects in one image) and slice it with python3 ${ASSET_GEN_SKILL_DIR}/tools/grid_slice.py kit.png -o out/ --grid 2x2 --names "a,b,c,d".
• Direction. Generators get left/right and facing wrong; generate one direction and flip at runtime.
• Audio format. Every tool picks it from the -o extension. Bevy's default features decode only Ogg Vorbis, so write .ogg for Bevy.
• Review before building on it. Look at every PNG. You can't look at a GLB or hear a WAV: render GLBs (${GODOGEN_ASSETS}/README.md, Checking results) and read audio as a spectrogram (${GODOGEN_ASSETS}/sfx/README.md).

Moves that drive gameplay — root-motion locomotion, attacks, state machines, anything held, ridden or aimed, characters that touch each other: read ${ASSET_GEN_SKILL_DIR}/motion.md.

Blender

Some 3D is better built than generated. A bpy script run headless (blender -b -P make_turret.py, exporting with bpy.ops.export_scene.gltf) is often the better path when:

• the shape must be exact or articulated: a vehicle whose wheels turn, a door on its hinge, a turret that aims, kit pieces that snap together;
• a generated GLB needs work: split into parts, new pivots, decimation, several merged into one;
• the animation is outside the humanoid tools: mechanical parts, a creature of a few bones, keyframed or procedural clips.

Check what it writes like any GLB. Without blender on PATH, asset-blender script.py -- args runs the script in godogen_assets' Blender (its print goes to stderr).

Paid APIs

${ASSET_GEN_SKILL_DIR}/api.md covers Gemini / xAI Grok images, animated sprites from Grok video, Tripo 3D, Gemini TTS and Lyria music. Read it only when an asset needs one:

• godogen_assets isn't on this machine (its checkout above reads none), or a command answers not installed — that part isn't set up;
• the local tools don't make it: frame-by-frame animated 2D sprites, music, one character voice acting many emotions, vocal barks, two-speaker dialogue, a language qwen-tts lacks, a rig for a non-humanoid;
• a local result stays wrong after rewording and a few seeds.

Paid calls cost real money: confirm with the user before the first one.

Asset manifest (in README.md)

Track every generated asset in README.md with an in-game Size column — without it, coders consistently scale assets wrong. gen3d models come out 1.0 tall with no real scale, so this column is where their size is decided:

• 3D models: meters, e.g. 4m long, 1.8m tall, 0.3m
• Textures: tile size, e.g. 2m tile
• Backgrounds: pixel size + behavior, e.g. 1920x1080, fullscreen
• Sprites: display pixels, e.g. 128x128 px
• Audio: length + use, e.g. 1.2 s one-shot, 8 bars / 22.6 s loop