I had used Flux.1 for couple of days and I had trained 3 loras using ai-toolkit, it performs better than SD series models, I don't doubt that. But think about that, it is so huge! Better generating quality is the thing it should done. Over 20G file size makes it hard to training even hard to use, what I expected is that it will bring some brand new features on generating (I'm not talking about text generation). But so big a file? OK, I guess it should be bigger to realizing it's potential...but where is the potential?
The cover image was generated by Flux.1 with a lora trained with 1250 steps on my soles dataset for test, feet fused together as before like SD1.5 or SDXL, you may not know such things if you don't care about feet or hands closeups. The face, usually the best one object all models generate great, will not be so hard to training, faces won't melted together, there won't be 3 eyes on one face, the mouth can't be on the top of head...Whatever object you train, if the model itself can handle something similar, it learns your concept, but the things model don't know, that's the problem, for example, the feet soles. I found that all the mistakes what SD models made, is now made by Flux.1, you may see some advantages like 5 fingers or good anatomy on Flux.1, but it does not perform as better as it's size. It should be much much better than what we saw before, right? If there was a 2B version of Flux.1, I doubt it could be better than SD3, assuming both with some community efforts.
SDXL generates images with better quality than SD1.5, but with larger size, with more difficulty of training, it was not as successful as SD1.5, there are still some new techs only suitable for SD1.5, even recently, like IC-Light. What about Flux.1? It's too big and too slow for individuals, without any unique features except text generation, I guess there will be a "Flux.2" or "Flux 1.0", with either unique generating ability or smaller size less than 10G, if not, time won't help it gain users.