I would call it Comics. Its the best at it so far out of all flux models.

Try safety_checker=None when loading the model and make sure you train with guidance scale of 1. Also I have noticed either t5 or CLIP has a problem with NSFW terms. I am pretty sure its t5 doing it because I tried using clip from NSFW SDXL models and it worked a little better but not great. I love this model though! Great work for reals. You should consider a SFW and NSFW version.

@XEZ Also, I would like to see a 4-8 step version. Keep up the good work!

@MEGACHAD I have never heard of a safety_checker before. What UI are you using ? I use either Forge or ComfyUI

@XEZ lol a Pussy-fix lora ? We are so far from PonyXL... haha maybe next year, if PonyXL dev finally give up on AuraFlow

@Tetsuoo I am loading it with python. Comfy and Forge are too limited so I use my own code.

@XEZ @Tetsuoo I just checked and looks like safety_checker has been removed so it no longer applies I guess. I think your problems might be from the t5_xxl not being trained on NSFW content? You could try using a known NSFW CLIP tokenizer/encoder to see if that will work. If the keywords are not tokenized it won't be able to learn NSFW. If that doesn't work then its probably because text_encoder_2 t5_xxl needs better training.

however I use stable diffusion, and flux doesn't work in my opinion

Images are all pixels.

@ApexThunder_Ai I went into this problem too, in my case the text encoder did not match the model that I was using, so just changed it, by the way, I'm using GTX 1050

Sorry, I am not sure if you figured it out by now since it's been 3 months, but here you go anyway!

Encoders for FLUX: AE + Clip L + T5 FP16 or T5 FP8 (lighter).

Clip L & T5 goes into webui\models\text_encoder
VAEs goes into webui\models\VAE

@Hyokkuda I've already found the solution. But thanks for help anyway!