doggel @[email protected]

@thor Almost forgot, nowadays, every single model out there is a merge of some damaged model, if you want to get something good and update, you may have to do weighted u-net layer merging of older models with new models and pray that it comes all right or at the very least you get more insight in what exactly you are doing.

@thor Lose of information also gets passed down when merging.

@thor (2/2) re-train a working model with not-properly tagged pictures of corgis, now all dogs are going to start looking a little more like corgis

@thor (1/2) That depends heavily on what model you are using, most models out there are merges, when you merge a model, defects in one model get passed down. A very common defect is ignoring the tokens in the prompt, like for example, ignoring the second token, or the fifth. This is lowkey intentional. Also, moving up the cfg scale a little will also help. Another problem is that retraining models make the model lose info, for example, if you...

@thor (2/2) In the negative prompt "more than 2 hands", the model will find it hard to generate good hands, since it is both being instructed to generate hands (2) and to not generate hands, and pictures with 2 hands and pictures with more than 2 hands share way too many similarities (the presence of hands). This will result in an increased rate of deformities in the hands. To properly add negative prompt is best to keep things short and precise, or use embeddings.

@thor (1/2) Simply put, if you type "anime girl" in the prompt, the diffusion model will generate something, anything, that could fit the "anime girl" tag, that may include missing limbs or deformities. Because of this, you could either add "cute" in the prompt or "zombie" in the negative prompt and it will try to generate something that fits those characteristics.
With that been said, blindly adding negative prompts will backfire, let's say you put in the prompt "2 hands" and...