diversity of outputs suffers if 0.24_dmd2 is increased to a high total value (e.g. 1.3x total)

however, you can sample with this (0.9/0.26/0.42) setup for 7/12 steps with euler_a, then start at step 6 and finish with an additional +0.8x of 0.24_dmd2 to gain most of the clarity benefits without the diversity penalty

bonus points if you further refine the output with a start at step 11/12, with add_noise disabled

You don't need to use heun and beta for that. Just add "realistic" to the prompt.

@rift1000 I like the skin texture more on heun. By far.

You're not transforming anything. The anime model is based on a realistic model; it's an attempt to retrain the anime to make it anime. Because the base model has incredible spatial awareness, which allows it to do all sorts of things. The base model is truly unique, but few realize just how unique it is compared to popular models, which don't even come close in terms of capabilities.

@sizyking I don't really care for cosmos predict realism capabilities. I can mix realism with the anime knowledge, which was interesting to me, specifically character knowledge.

@deitychaser
It's not about knowledge of the world, but rather that it allows the model to generate a user query. A typical model generates only what it has specifically learned. This model simulates the query, meaning it can generate something it hasn't been specifically trained to generate. It loses to models that claim to be realistic, but they're only aesthetic realism. This model simulates realism, which is worse than the natural ones. Some FLuX displays the aesthetics of real photos, and people see quality in this, but FLuX doesn't understand what it's generating. This model understands, and that's what makes it unique.

@sizyking Sorry, but this is just wrong. The reason Anima still does realism well is because tdrussel trained mainly the LLM adapter and the actual backbone had seen so far only little training. That's the reason why if you replace the text encoder with just t5 then Anima will just revert to cosmos knowledge entirely and forget almost all of the training it had so far. You can test it out, its a generally known quirk was the model, although he started shortly after the release of preview2 to freeze the adapter and train the backbone only, so with future releases Anima will probably become worse at realism. Another factor why it specifically works well with this distill lora is because the creator is known for having tons of photography in his datasets.

@deitychaser I tried T5, it draws anime too, but it needs to be described specifically because T5 is just a translator, and it doesn't have the logic of Qwen. I only achieve realism with natural language; if the model sees tags, it immediately wants anime. I tried T5, Qwen0.6, Qwen4b, they all make anime if I write tags. But Qwen is still better than the basic T5 because it has logic, but less knowledge.