aziibpixelmix-onnx

AziibPixelMix-ONNX

Browser/WebGPU-ready ONNX export of AziibPixelMix Fast Mode (pixel-art SD1.5 + LCM): dynamic UNet (fp16), clip-skip-2 CLIP (fp32, final-normed), dynamic VAE (fp32), DPM++ SDE timetable. First load downloads ~2.4GB once, then cached; 4-step images render fully on-device via ONNX Runtime WebGPU.

Credits (required by the creator's terms)

• Model: aziib — pixel-art SD1.5 mix. Full version also on Hugging Face: megaaziib/aziibpixelmix. Trigger words: pixel art, pixel world. Fast Mode recipe: 4–8 steps, CFG scale 1–2, DPM++ SDE sampler, clip skip 2.
• Base lineage: Stable Diffusion 1.5 (Realistic Vision V4.0) with LCM distillation.
• If you use the model commercially, the creator asks for support: ko-fi.com/megaaziib. Generated images are OK for commercial use per the creator's terms.

License & sharing note

Published here with the creator's permission. The weights follow the upstream Civitai terms (credit required); the runtime scripts (sd_web.js, sd_web.ts) are Apache-2.0.

Files

Reproduce the conversion

pip install torch diffusers transformers huggingface_hub onnx onnxruntime onnxconverter-common
python tools/convert_sd.py --ckpt aziibpixelmix_fastModeExperimental.safetensors --out-dir ./aziib-onnx
python tools/render_sd_onnx.py ./aziib-onnx "pixel art castle, pixel world" out.png --steps 4 --cfg 2 --seed 1 --w 512 --h 768

The script exports UNet (fp16, dynamic batch + size, parity-gated), CLIP (fp32 export, then post-hoc fp16 conversion, parity-gated), and VAE (fp32, bit-exact), then writes the scheduler timetable and config. tools/supra_arch.py is a shared helper it imports.

Run it in JavaScript (no build step)

Requirements: a WebGPU browser (Chrome/Edge 113+ with hardware acceleration) served over http(s) — file:// will not work.

<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/ort.all.min.js"></script>
<script type="module">
  import { loadSD, generateSD, paintSD } from './sd_web.js';
  const model = await loadSD('https://huggingface.co/Bartholomheow/aziibpixelmix-onnx/resolve/main/', { onProgress: console.log });
  const { pixels, width, height } = await generateSD(model, 'pixel art castle, pixel world', { seed: 1 });
  paintSD(pixels, width, height, document.querySelector('canvas'));
</script>

Run it in TypeScript (npm project)

npm i onnxruntime-web @huggingface/transformers

import { loadSD, generateSD, paintSD } from './sd_web';
const model = await loadSD('https://huggingface.co/Bartholomheow/aziibpixelmix-onnx/resolve/main/');
const { pixels, width, height } = await generateSD(model, 'pixel art castle, pixel world',
  { seed: 1, steps: 4, cfg: 2, width: 512, height: 768 });
paintSD(pixels, width, height, document.querySelector('canvas')!);

Options: { seed, steps, cfg, width, height, onProgress } (onProgress reports { phase: 'encode' | 'denoise' | 'decode', step, steps }). Defaults (4 steps / CFG 2) match the creator's Fast Mode recipe; any 512–1024 resolution works.

Troubleshooting

• No available backend found → WebGPU needs hardware acceleration (Chrome/Edge Settings → System). VMs, remote desktops, and old drivers often lack it.
• First load stalls → ~2GB download, then cached. Weak iGPUs still work, just slower.
• CPU-only browsers: the 1.7GB fp16 UNet needs WebGPU-class memory and will not initialize on small WASM heaps (std::bad_alloc at session creation) — WebGPU is effectively required for the UNet; CLIP/VAE alone do run on WASM. Sessions are created sequentially to keep peak memory as low as possible.
• Black image with no error → NaN poisoning (NaN paints as black). Both runtimes fail fast with text encoder returned NaN… instead of rendering black.