Flux.dev as we have known until now was a distilled model, meaning it was trained by flux.pro as its teacher. These new models change everything ! This is the first experimental, effectively de-distilled version of Flux.dev, meaning it is much closer to what flux.pro is capable of. And it's just the start ! (NB : This is not stating it was trained with flux.pro - I don't know the exact method) PARAMETERS You can now forget about Distilled CFG and use real CFG (I've tried up to like 14) NEVER, ever use it with CFG = 1 - It will automatically be a complete disaster, and most of the time the reason why you don't get results You SHOULD use at least 40 - 60 steps, depending on the CFG you use. It will be much longer but SO worth it Unfortunately the current hyperdev 8-steps Lora doesn't seem to work with it to reduce steps Dedistilled allows NEGATIVE PROMPTS BENEFITS Prompt adherence will be EXCEPTIONAL, even with Loras. Faces Loras will work better, details will be better, text will be WAY better... Everything from the prompt will be better basically