Quality Assurance Program
The quality of the output of the 4-bit model can be effectively maintained by the following technical means:
- SVD outlier handling::
- Automatic identification of outliers in weight matrices
- Preserving eigenvalues above 95% by low-rank decomposition
- Mixed accuracy compensation::
- Use FP16 calculations for the Attention layer (needs to be set up)
attn_precision=fp16) - The VAE decoder forces the use of FP32 (via the
force_full_precision_vae=True)
- Use FP16 calculations for the Attention layer (needs to be set up)
- Post-processing optimization::
- Load ADetailer extension to automatically fix facial details
- Super-resolution reconstruction with TileDiffusion
Real-world comparison data:
- On the FLUX.1-schnell model, the FID score for the 4-bit quantized version is 18.7 vs. 17.2 for the original version
- Human reviews show that slight color banding is only observed when zooming in on the 400%
Operational Recommendations:
1. For critical items, first generate 256 x 256 sketches
2. Adoptionhq_upscale=2.0Parameters for 2x superscoring
3. Finally, apply Nunchaku's ownDebandFilternodal
This answer comes from the articleNunchaku: an inference tool for efficiently running FLUX.1 and SANA 4-bit quantization modelsThe
May not be reproduced without permission:AI productivity tools " How can I avoid the degradation of the quality of the generated image due to 4-bit quantization?
English 
简体中文 
日本語 
Deutsch 
Português do Brasil