RDBT [Anima]
A general finetuned model, better overall quality, better prompt adherence, less artifacts and errors.
This page contains LoRA files for advanced users.
See the main page for more info.
See this page for update log and version info.
FAQ:
This LoRA can't remove nor resist Al slop effects.
If you got overwhelming Al slop effect. Read more here.
If I didn't release the LoRA for new version, means
Update is small, and I don't know if it works.
I trained the new version as accumulated LoRA, Then I have a LoRA strength bar for new data/method. Then I can easily test it.
However releasing multiple accumulated LoRAs for one version is quite tedious and misleading for most users. So I just release a merged ckpt.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Special trigger words and latent watermark included.
Description
FAQ
Comments (8)
Thank you for your efforts! Descriptions says Workflow included, but can't find it. Can you please share?
Excellent works.
After experimenting a lot with your loras, i decided to take a look at the base model again.
While slow as molasses and more prone to errors, it reminded me of how much more prevalent the diversity and style adherence are in relation to any of your distillations (both dmd2 and regular cfg distill). It's to be expected, given the process, but is there a non-distill lora in your plans, or is it at least feasible to approach that level of style adherence in a distill?
finetuned-only version: No, releasing 3 versions is very confusing. You can basically think of the cfg distilled version as "no-distilled" version, if you use it with cfg. It might change the output, that's all trained models will do. But it does not let model forget things. Update: it does forget things, "bad quality, score_1,username...", the things that in negative prompt template.
approach that level of style adherence in a distill, this is the problem of dmd2, which model needs to forget things, that's how dmd2 works: I would say it's possible, but impossible to me. I tried, preserving top500 artist styles during distillation, and failed, and I estimated that I will need 10x+ compute to do it. And there are 10k+ styles.
I'm quite confident about the diversity of cfg distilled model.
p3 v0.24f dmd2 doesn't work properly for me with the same parameters as version 0.23. I tried steps 6, 12, and 16. The results are poor. How do I use it?