Instructions to use circlestone-labs/Anima with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusion Single File
How to use circlestone-labs/Anima with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Optimal sampler / scheduler settings for Anima (+ comparison site)
Hi everyone β I spent some time testing samplers, schedulers, and hires fix settings on Anima, and I wanted to share what I found. I also built a small site where you can compare the outputs side by side:
π https://anima-compare-viewer.pages.dev/
res_multistep + beta: fast, but I sometimes get leftover dot/polka-dot artifacts.
ipndm_v and kl_optimal: often fail to produce a clean output for me.
After comparing across a range of prompts, the fastest combination that also gives consistently clean results was uni_pc + ddim_uniform. That's what I'd recommend as the default for Anima.
Hires fix
Anima seems to cap out at around a 1.5Γ upscale in a single hires fix pass. Pushing it higher in one go tends to break down. What works best for me is running hires fix twice: the first pass at 1.5Γ, and on the second pass I specify the target resolution directly (2160 in my case) instead of using a multiplier. If you're happy with a smaller output, you can of course skip this and stay at the base size.
Samplers / schedulers
You can see the side-by-side comparisons here: https://anima-compare-viewer.pages.dev/
Curious if others are seeing the same thing, or if there's a combo I missed. Happy to add more samplers to the comparison site if there's interest.
This is very nice, but I highly recommend redoing the CFG++ samplers at a lower CFG. You're supposed to use them at a CFG below 2 (I recommend 1.5), at 4.5 they're going to output heavily fried images or just garbage in general which you can see in the outputs
For the site it's annoying that I can't see the full prompt, I would want a clickable text that copies the prompt to the clipboard. Same for negative prompt.
As mentioned a lot of outputs are fried. You can use https://github.com/Extraltodeus/Skimmed_CFG to avoid this. I don't have the full prompt/negative prompt but will post some examples. Top pictures are from the site.
Prompt: 1boy, short blonde hair, green eyes, knight armor, castle hallway, holding shield, holding sword,
Negative prompt: (worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, monochrome, watermark:0.8)
Euler CFG++
Euler CFG++ (skimming at 1.5)
Euler Ancestral CFG++
Euler Ancestral CFG++ (skimming at 1.5)
You can also drag my outputs into ComfyUI to get the same result as the workflow is embedded.
Also not sure what "cleanest" result means with uni_pc + ddim_uniform, that combination for me gave worse results than those I posted. uni_pc with simple looks better. Same with fastest, right now you're comparing everything at 30 steps while the CFG++ samplers and Seeds 2 and a lot of other samplers look good(ish) at 16. I think the best sampler should be something that looks good and also follows the prompt.
The regular Skimming node works right?
That looks amazing, I put off anima due to fried outputs, now maybe it will be amazing!

