I come to terms that it's maybe not impossible, but most certainly too much hassle to make something work like i want it too. After a lot of custom scripts, trying different approaches, text encoder training , changing how trainers actually train models, to a point where i made a lora that should not work but it did... Training just a few images that have a constant theme, like photorealistic, is no problem. A big dataset which maybe has the same theme, but the variation in style are so extreme, require a lot of steps and care with the learning rate, resolutions, sometimes captations are to "accurate" which can lead to early overfitting ( streaks, artifacts ) not enough time to learn the picture before collapse, yadayada. Baseline is, Chroma still suffers from the model it was trained on, even now. Low Steps, Destillation ( Flux Schnell ) and maybe even locked weights are absolute poison for training, because that stuff restricts, makes a model rigged and narrows down the variations. Also a reason why i stay away from Z-Image for now till they release the base. But even then it's Alibaba with the good ol' chinese censorship and tons of concepts it doesn't even know, which most likely requires a few months of fine tuning to get a real "Base". Really hope that happens, because it's more light-weight, good quality and has a very good understanding of prompts. Who knows. Maybe more my own fault then the model itself, trying to brute force something that in general is a bad approach to training... just a reason to do it anyway :)

Files with SHA256: ee66c3735253f3afc3f5b0db41138b773a26ac8c28d858715e64f1bb8d105290