This is the result of what I call "phase 1" training for the model at https://huggingface.co/opendiffusionai/XLSD-V0.1 (A merge of the SDXL VAE, and the SD1.5 unet)
The model has been retrained to the point where at least the colors are sane. But it still has a lot of work to undergo, before it looks like a NICE model.
For the curious, this was trained on 165,000 images from the CC12M dataset. Training settings were:
BF16 precision. Prodigy optimiser, batchsize 8, Cosine with hard resets(10 cycles). No EMA, 1 epoch
Sample image below. As one can see, the colors are now reasonable, but... objects are fuzzy. There is still a significant amount of retraining to be done. This model probably needs the last (10% ?) of training typically done when training a model from scratch.
Also, I dont know what I'm doing, so I may need to just redo phase1 completely anyway. But if anyone wants to play with the current incarnation: here it is!
