FLUX + FLAN

Does quantizing from FP32 really make a difference? I would argue that it does based on the results I have gotten from the latest model.

The UNET is untouched, no lora merges or other alterations. This is important as a balanced finetune is near impossible on FLUX as many have noted.

How is FP32 UNET possible then. Over one trillion calculations where done (Monte Carlo) to predict the model to FP32 - This is the one step or 0.5% accuracy estimation

10,000 ~0.005 (0.5%) 2 digits
1,000,000 ~0.0005 (0.05%) 3–4 digits
1,000,000,000 ~0.000016 (0.0016%) 5 digits

FP32 TE is Google FLAN pruned, CLIP-L is Zer0int FP32 Detail CLIP

NF4 is extremely quick and with Forge you can pair it with LORA's - I would use NF4 over GGUF or FP8 in any instance when quantizing