In all fairness, it's the first Flux 3 Lora on this site, and it says it's a test. I'm training a test myself right now, and it's showing some promise. We'll need a lot of "awful" attempts before we all figure out how to make the good ones.

@Kuminer69 i think he is talking about the model itself... its almost as bad as mage flow but atleast this one gives a output back

@juserisesseries580 I guess I look at it as, this is probably about as good as we're going to get for a model that can run on 12GB VRAM, and also be trained on the same. I see some of these absolutely insane image/video models, and they're just not realistic without an insane pc, or renting a GPU.

@Kuminer69 i am trying to make it work with krea2, so krea2 edit, into flux3 as "face swap" model, i might try refmod for face swapping, i tried to get antigravity cli to do it but they have removed all the pro models as of now, only gemini 3.8 flash remains, and argon gemini 4 is for like non promotional paid plans(so stupid 3.1 pro was dumb at best), so might have to do it manually i am using int4 krea2 turbo model, 2k-4k in like 30to90seconds with really good generations with a lot of lora support. on 8GB VRAM BTW and 16GB RAM.

edit: yes video models are basically non existent atleast if you want grok like fast inference(2 min and under), minimax h3 works tho if you are patient enough.

@juserisesseries580 That's cool. I don't think Lora training on Krea2 is going to be possible on my machine though. I'm more interested in experimenting with training, than I am with finding the model with the most realistic generations, so I'm not sure that workflow would be suitable for me, but I'm ecstatic that it works for you. I'm just not a big fan of leaving comments on someone's Lora, like OP did on this one, just to diss the model they are training on. My only real reason for commenting on this in the first place.

@zoefollonier Nice. Keep up the good work.

@zoefollonier Im really interested in your tools for training loras with krea2. Sitting on 16 GB vram

@Kuminer69 it's not about your job. Sorry