hanasaan/frag-loop

Frag Loop

Frag Loop is a decoder-only Transformer trained to generate Shadertoy-compatible GLSL fragment shader bodies. The model outputs the body only (functions + mainImage), and is intended to be wrapped by a fixed WebGL2/Shadertoy template at runtime.

Intended Use

• Primary: generate novel GLSL fragment shader bodies compatible with Shadertoy-style templates.
• Not intended for: general chat, safety-critical systems, or production code generation without review.

Training Data (Summary)

• Pretraining: GLSL-heavy subset of The Stack (bigcode/the-stack-dedup).
• Fine-tuning (SFT): Vipitis/Shadereval-inputs (Shadertoy-oriented subset).

These datasets contain code under multiple licenses. If you redistribute outputs or data, ensure provenance and license compliance.

Tokenizer

The tokenizer implementation is heavily influenced by nanochat:

• https://github.com/karpathy/nanochat

Training Procedure (High-level)

• Decoder-only Transformer (GPT-style) trained from scratch.
• Pretrain on GLSL-heavy corpus, then SFT on Shadertoy-style examples.
• Optional copy-detection and automatic rejection are supported in the runtime.

Evaluation

• Primary evaluation is compile + render for WebGL2 (moderngl), with metrics for:
  • compile success
  • render time
  • black / static detection
  • NaN / Inf detection

How to Use (Recommended Runtime)

The reference runtime + WebUI are provided in the GitHub repo:

• https://github.com/hanasaan/FragLoop

From the publish/ bundle, you can run:

# download from HF and start inference server
python infer/inference_server.py --hf-repo hanasaan/frag-loop

# start UI server
node ui_optional/server_node/server.js

Open http://localhost:5173 in a browser.

Limitations

• Outputs may occasionally fail to compile or render.
• The model can still produce outputs similar to training data. Use copy-detection when displaying or distributing outputs.
• Shader aesthetics and performance vary; manual curation is recommended for showcases.

Acknowledgements

• Tokenizer influenced by nanochat (https://github.com/karpathy/nanochat).
• All implementation work in this repository was done by GPT-5.2-Codex (xhigh).

License

This model is trained on datasets with mixed licenses. Please review and comply with dataset licensing terms when using or redistributing the model or outputs.