RDBT [Anima]

Finetuned circlestone-labs/Anima.

This page contains the ckpt version, for convenience.

RDBT was trained as a LoRA for better training and distribution efficiency. Main model page: https://civitai.com/models/2364703.

File reupload.

Turbo (4-step dmd2)

Fast as f*** boiiiii

4-step dmd2, distilled on top of RDBT finetuned base model. See the link above.

Settings:

• Sampler: "euler_a" or "euler".

• CFG scale: 1.

• Steps: 4, 8 (recommended), or 16. Scheduler: simple. Important: training timestamps are fixed. Other inference timestamps might not work.

First time doing 4-step dmd2, also first 4-step dmd2 anima model. I don't know what I'm doing and what to expect... Highly experimental.

I was curious about what a 4-step dmd2 anime model would look like. And here it is.

Huge stability improvement, it even can render longer text in 4 steps.

But pushing model to 4-step might be too much, and tbh, not necessary. Because nobody really uses 4 steps, because 4 steps is doable, but unstable, there are lots of small errors. Most people use 8 steps.

If you want to compare:

8-step dmd2 https://civitai.red/models/2364703?modelVersionId=2832699

16-step dmd2 https://civitai.red/models/2364703?modelVersionId=2860424

I think I will stick with 8-step next time. Open for suggestions.

===================

p1 [fp8]:

fp8 anima preview. Uploaded here for convenience.

About fp8 models:

Quantized fp8 models in ComfyUI format. If you GPU supports it (aka. rtx 4000 and later), ComfyUI will use hardware fp8 automatically, which might (blame Nvidia) be faster. More about ComfyUI fp8, see ComfyUI docs TensorCoreFP8Layout.