AssetForge

A local, offline pipeline for generating 2D game assets: sprites, tilesets, icons, characters, sound effects and music. Everything runs on your own GPU. Nothing is sent anywhere.

Three engines share one repository:

• Image drives ComfyUI headlessly across Flux Schnell, SDXL and SD 1.5, then post-processes renders into real pixel art with correct grids and palettes.
• Audio drives Meta's AudioCraft (MusicGen and AudioGen), then post-processes raw output into trimmed, level-matched, seamlessly looping game audio.
• Mesh turns one keyed render into a textured GLB with Microsoft's TRELLIS.2 and renders it into 2.5D sprite frames with Blender, so a turret or hull is consistent from every angle.

The generation half is the easy half. Most of what is here is post-processing, because a raw diffusion render is a picture of a sprite and a raw AudioCraft clip is a recording that resembles a sound effect. Turning either into an asset an engine can use is the work.

Requirements

• NVIDIA GPU (developed on a 16GB RTX 4060 Ti)
• Windows, Python 3.11 for the audio engine
• About 40GB of disk for models

Install

Neither engine's runtime is in git. docs/SETUP.md walks through both from a fresh clone, including the exact dependency pins the audio engine needs and a troubleshooting table.

Short version, once installed:

python engines/image/scripts/download_models.py --all
cd engines/image && ./python_embeded/python.exe -s ComfyUI/main.py --windows-standalone-build --listen 127.0.0.1 --port 8188

Then, from the repository root:

python engines/image/scripts/asset_presets.py --preset item_icon --subject "a red health potion" --output out/image/potion.png

For audio, use the audio engine's own virtualenv, which pins an older torch than the image engine:

./engines/audio/venv/Scripts/python.exe engines/audio/scripts/audio_presets.py --preset sword_hit --output out/audio/sword.wav

Both asset_presets.py and audio_presets.py take --list.

What it produces

Asset	Command
Item icon, true 32px indexed art	asset_presets.py --preset item_icon
Side-on character sprite	asset_presets.py --preset character
Seamless terrain tile	asset_presets.py --preset topdown_tile
A full tileset	gen_tileset.py then make_tiles.py
A composed map	build_village.py
Engine-ready terrain transitions	autotile.py (47-tile blob atlas + JSON)
Normal maps for 2D lighting	make_normalmap.py
Sound effect	audio_presets.py --preset sword_hit
Looping music track	audio_presets.py --preset town_theme
Clean alpha for a rendered sprite	remove_background.py (BiRefNet)
Same object from another camera angle	edit_asset.py --model qwen-edit --angle ...
Textured 3D mesh from a render	engines/mesh/scripts/generate_mesh.py
Sprite frames from a mesh	engines/mesh/scripts/render_sprites.py

Demo

demos/moonlit_ruins is a complete vertical-slice pack for a platformer: 11 images, 10 audio files, and a scene composed from them. Nothing in it was drawn, recorded or edited by hand.

Documentation

• docs/SETUP.md — install both engines, and troubleshooting
• docs/RECIPES.md — worked end-to-end workflows
• docs/IMAGE.md — image engine, post-processing, tilesets
• docs/AUDIO.md — audio engine, looping, game audio conventions
• docs/MODELS.md — every model, when to use which, measured comparisons
• docs/FINDINGS.md — what was measured, and what was tried and rejected
• CLAUDE.md — condensed decision rules for an agent driving these tools

docs/FINDINGS.md is the one to read before changing anything. Several obvious improvements were tested here and measured to be neutral or harmful, and the reasons are specific to this stack rather than general.

Layout

engines/
  image/
    ComfyUI/          runtime (not in git)
    python_embeded/   runtime (not in git)
    workflows/        one ComfyUI API workflow per model
    models.py         image model registry
    scripts/          generation and post-processing
  audio/
    venv/             runtime (not in git)
    scripts/          generation and post-processing
  mesh/
    venv/             Python 3.11 + torch 2.7 (not in git)
    ComfyUI/          second ComfyUI, port 8189 (not in git)
    ComfyUI-Trellis2/ TRELLIS.2 node with its wheels (not in git)
    scripts/          image to GLB, GLB to sprites
out/                  generated assets (not in git)
vendor/               third-party sources, see licences below
docs/

Licences and attribution

The pixel-art grid snapper in engines/image/scripts/pixel_art_processor.py is vendored from doriansao/codex-pixel-art-generator (MIT), itself a port of Hugo Duprez's spritefusion-pixel-snapper (MIT).

Model weights are downloaded from Hugging Face under their own licences, which differ. FLUX.1-schnell is Apache 2.0, SDXL is CreativeML Open RAIL++-M, and AudioCraft's weights are CC-BY-NC 4.0, which is non-commercial. Check the licence of any model before shipping what it generates.