shader-forge

An autonomous shader studio — a local LLM writes WGSL/GLSL art shaders, naga + 6-axis LLM review gatekeep them, and a human curator (you) approves the final gallery.

Status

Research prototype, currently dormant. Phase 3 (WGSL quality lift) paused mid-flight while the author focuses on another project. Code, gallery (55 shaders), and design docs are preserved as-is. Issues welcome for reproduction questions; PRs deprioritized until the project resumes.

Why it exists

Standalone WGSL art shaders barely exist in the wild — most WGSL corpora are engine internals (Bevy PBR, wgpu examples, Web-RTRT pipelines). shader-forge was an experiment to see whether a local 80B code model plus a curated RAG could produce something closer to Shadertoy-quality output, but in the WebGPU generation.

It's also a concrete reference for fully local, autonomous creative loops: no OpenAI call, no hosted inference, no external review API. Everything runs on one box.

Pipeline

SCOUT    → Tavily API pulls trends → LLM proposes 3–5 shader themes
FORGE    → RAG selects references → Qwen3-Coder-Next 80B writes WGSL/GLSL
           → naga-cli validates → LLM reviews on 6 axes (30 pts total)
           → on failure, feedback-driven retry (max 5×)
RENDER   → headless WebGPU (Chrome + Vulkan) → thumbnail.png + preview.mp4
STAGE    → FastAPI WebUI on :3333, dual WebGL/WebGPU live preview
PUBLISH  → approved shaders drop into inbox/ for downstream pickup

Stack

Layer	Tech
Inference	llama.cpp / llama-server on localhost:8080
Model	Qwen3-Coder-Next 80B-A3B (Q4_K_XL, ~47GB)
Shader compiler	naga-cli (WGSL → SPIR-V), glslang (GLSL ES)
RAG	ChromaDB + all-MiniLM-L6-v2
Headless render	wgpu-py + Vulkan, Puppeteer + Chrome headless
Review WebUI	FastAPI + Three.js (WebGL) / raw WebGPU

Designed for NVIDIA Project DIGITS (GB10 / ARM64), but any CUDA + Vulkan host with enough VRAM for the model should work.

Highlights

• 514-shader curated WGSL corpus from wgpu/Bevy/compute.toys/approved art + 23 shabon-fx GLSL references → 96.1% passed naga validation
• let→var auto-fix for the most frequent WGSL-vs-GLSL translation error the LLM makes
• 6-axis review rubric (composition / motion / color / novelty / code quality / concept fit) — JSON-parsed, feedback fed back into the retry loop
• Best score shipped: 24/25 (vibrant-plasma-energy-vortex)
• Temperature split: creative director prompt runs at 0.7, the coder stays at 0.0 — separation of design intent and faithful implementation

Gallery

55 generated shaders live in gallery/, each folder carrying:

• shader.frag (GLSL) or shader.wgsl
• metadata.json — theme, style hints, references used, review status
• thumbnail.png — 1080×1080 render

Browse locally via the Stage WebUI:

cd stage && python -m server  # http://localhost:3333

Quick start

# 1. Environment
cp .env.example .env
# edit .env: set TAVILY_API_KEY and point LLM_BASE_URL at your llama-server

# 2. Python deps
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 3. naga-cli
cargo install naga-cli

# 4. Start your local llama-server (example)
llama-server -m /path/to/Qwen3-Coder-Next-UD-Q4_K_XL.gguf \
  --port 8080 --host 0.0.0.0 -ngl 99 --ctx-size 16384 --flash-attn on

# 5. Generate a WGSL shader
python forge/cli.py --theme "aurora borealis" --target wgsl
# Output lands in gallery/YYYY-MM-DD_aurora-borealis/

# 6. Review WebUI
python stage/server.py  # open http://localhost:3333

Not included in this repo

• context/ — the RAG corpus (third-party Shadertoy/Bevy/wgpu code). Regenerate via forge/fetch_shadertoy.py and forge/extract_repos.py.
• data/chromadb/, data/qlora_output/, data/training/ — reproducible build artifacts.
• Web-RTRT/, WebGPU-Fluid-Simulation/, WebGPU-Lab/, webgpu-demo/, webgpu-native-examples/ — upstream reference projects; clone them separately if you want those samples available to the ingest pipeline.
• Model weights — download from Hugging Face.

Design docs

• DESIGN.md — v1 (WebGL1 / GLSL ES 1.00) architecture
• DESIGN_v2.md — v2 migration to WebGPU / WGSL
• CHANGELOG.md — running log of experiments and key findings
• WEBGPU_SHADER_REPOS.md — survey of WGSL corpora

License

MIT

Links

• @wrennly_dev — updates
• shabon-fx — 24 curated GLSL shaders that seeded the RAG corpus
• react-shadertoy — the renderer used for live preview

---

日本語メモ

ローカル LLM だけでシェーダー制作パイプラインを自律化する実験プロジェクト。trend スカウト → RAG 参照選択 → WGSL/GLSL 生成 → naga バリデーション → 6 軸 LLM レビュー → headless レンダリング → 人間キュレーションまで、外部 API を (Tavily を除き) 使わずに完結させた。

WebGPU 移行 (Phase 3) の途中で別プロジェクトに集中するため休眠中だが、standalone な WGSL アートシェーダーという空白領域 に踏み込んだ記録として、設計書・コード・55作品のギャラリーを残している。