This is a Flux.1 dev full finetune model

use nearly 30000 image in 120+a100 hours tuned

This is still an early attempt for me to make large-scale adjustments to the FLUX model

Obviously, it seems that we have achieved quite good results this time

Please note that when using the model, please use CFG parameters and add negative prompts

Fine tuning seems to have reconstructed some CFG

Please control CFG at 3.5 for optimal results

The model has been integrated into FP8's checkpoint, and can be directly loaded using the load checkpoint without the need to separately load clips and t5xxl

I think these parameters are distilled parameters of the model. Should they be retained or removed in large-scale training

I think it needs to be removed. Experienced friends are welcome to reply and discuss

So, let's go back in time

enjoy it!