Sign In

FP32 (2GB-Q8) Refiners work amazing at 2k for FLUX or SDXL

8
FP32 (2GB-Q8) Refiners work amazing at 2k for FLUX or SDXL

FP32 Refiner

I was experimenting with using the FP32 refiner in FLUX and I am sharing my settings and findings.

The short of this is they work amazing but need a NSFW Finetune

Image of a Bird (FLUX)

Image of a bottle (SDXL)

In both images BRSGAN 2x was used to upscale a 1024x1024 image to 2048x2048

For the comparison the 2048x2048 image was down scaled using linear down scaling and the difference was overlayed.

You can see the changes are amazing. (The zip contains both the 1024 and 2048 images to compare.

1024x1024 VS the Refined Image

Note this also tends to beat the latest AI detection

Post with the original AI image

Impact Statement

  • Refiners work amazing well even when quantized to Q8.

  • They can be used even out of the normal scope with correct settings.

  • SDXL refiner does not work well on NSFW or even R rated images.

  • We need a NSFW finetune of FP32 SDXL refiner

8

Comments