This isn't my most stable training, but I might come back to this topic and try to improve the results. While this model isn't very stable yet, it can still produce decent results. So, I decided to upload it and let you experiment. Currently, the most stable sampler and scheduler combination for me is dpmpp_2m_sde and sgm_uniform. I trained with a trigger word, but it seems to have a negative impact on the final inference. You can also try g1ssm, but I didn't get any results with my model.

Files with SHA256: 4bca961fc0425c7d77172451302405497bd1944c1afbae58b2e34b6de62465fa