<h1 id="flux.1-dev-convrot-for-comfyui">FLUX.1-dev ConvRot for ComfyUI</h1>Native ConvRot quantized FLUX.1-dev models for ComfyUI. Use the W8A8 model for a balanced size and quality option, or test the other included ConvRot three diffusion-model variants and an optional INT8 ConvRot T5-XXL text encoder. <ul><li>Whole W8A8 + INT8 T5<ul><li>Global PSNR vs BF16: 26.559 dB</li><li>Mean Per-Pair PSNR: 28.872 dB</li><li>Low-VRAM Peak: 16,265 MiB</li></ul></li><li>Whole W8A8 + BF16 T5<ul><li>Global PSNR vs BF16: 27.399 dB</li><li>Mean Per-Pair PSNR: 29.449 dB</li><li>Low-VRAM Peak: 16,298 MiB</li></ul></li><li>Partial INT8 + BF16 T5<ul><li>Global PSNR vs BF16: 27.888 dB</li><li>Mean Per-Pair PSNR: 29.857 dB</li><li>Low-VRAM Peak: 20,351 MiB</li></ul></li><li>Paper W4A4 + BF16 T5<ul><li>Global PSNR vs BF16: 17.926 dB</li><li>Mean Per-Pair PSNR: 19.041 dB</li><li>Low-VRAM Peak: 10,797 MiB </li></ul></li></ul>Workflow:<a target="_blank" rel="ugc" href="https://github.com/kgonia/comfy-workflows/blob/master/FLUX1_Dev_ConvRot_API.json">https://github.com/kgonia/comfy-workflows/blob/master/FLUX1_Dev_ConvRot_API.json</a> Support my work: <a target="_blank" rel="ugc" href="https://ko-fi.com/michelangelofussion">https://ko-fi.com/michelangelofussion</a>Updates: <a target="_blank" rel="ugc" href="https://x.com/kgonia7">https://x.com/kgonia7</a>

FLUX.1-dev-ConvRot

t5 text encoder int8-convrot version

t5xxl_flux1_int8_convrot.safetensors

Files with SHA256: 78cd681f670b4bc064a8880e40bd4806aa266c6b484a2012846cbeae6f0bfe17