FLUX.1-dev ConvRot for ComfyUI

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.

• Whole W8A8 + INT8 T5

  • Global PSNR vs BF16: 26.559 dB

  • Mean Per-Pair PSNR: 28.872 dB

  • Low-VRAM Peak: 16,265 MiB

• Whole W8A8 + BF16 T5

  • Global PSNR vs BF16: 27.399 dB

  • Mean Per-Pair PSNR: 29.449 dB

  • Low-VRAM Peak: 16,298 MiB

• Partial INT8 + BF16 T5

  • Global PSNR vs BF16: 27.888 dB

  • Mean Per-Pair PSNR: 29.857 dB

  • Low-VRAM Peak: 20,351 MiB

• Paper W4A4 + BF16 T5

  • Global PSNR vs BF16: 17.926 dB

  • Mean Per-Pair PSNR: 19.041 dB

  • Low-VRAM Peak: 10,797 MiB

Workflow:

https://github.com/kgonia/comfy-workflows/blob/master/FLUX1_Dev_ConvRot_API.json
Support my work: https://ko-fi.com/michelangelofussion

Updates: https://x.com/kgonia7