How to run Anima locally.

Which version.

Anima is a 2B model from CircleStone Labs, made with Comfy Org and built on NVIDIA’s Cosmos-Predict2. It was trained on several million anime images and about 800,000 non-anime artworks, with no photos. All versions are the same size.

• Turbo (anima-turbo-v1.1) is distilled for 8 to 12 steps at CFG 1. The author recommends starting here: slightly behind Aesthetic on average, and much faster.
• Aesthetic (v1.1, or v1.0b without the merged style LoRAs) is the quality fine-tune for finished pictures.
• Base (anima-base-v1.0) has a plain, neutral style. Train LoRAs on this one.

Files you need.

Everything is in the model’s own repository. Only bf16 is official.

• Model anima-turbo-v1.1.safetensors ComfyUI/models/diffusion_models/ or anima-aesthetic-v1.1 / anima-base-v1.0, 4.2 GB each
  4.2 GB Download
• Text encoder qwen_3_06b_base.safetensors ComfyUI/models/text_encoders/
  1.2 GB Download
• VAE qwen_image_vae.safetensors ComfyUI/models/vae/ the Qwen-Image VAE. You may have it already
  0.3 GB Download

models/
├── diffusion_models/
│   └── anima-turbo-v1.1.safetensors
├── text_encoders/
│   └── qwen_3_06b_base.safetensors
└── vae/
    └── qwen_image_vae.safetensors

Smaller files: GGUF

Community GGUF builds load through the ComfyUI-GGUF Unet Loader. They matter for 4 GB cards; above that, the full file fits anyway.

The same uploader has Aesthetic and Turbo builds. Its page names the NVIDIA Cosmos licence; CircleStone’s licence applies.

What fits your computer.

At 1024 × 1024. Anima is small, but its attention makes it slower per image than SDXL at the same size.

• 4 GBSlow

  A GTX 970 runs the Q8_0 GGUF at under 7.5 s per step with fp16 compute.

• 6 GBFits

  The full bf16 model. Comfy’s template puts the whole set at 5.6 GB.

• 8 GBFits

  Full model with room for LoRAs.

• 12 GBFits

  Comfortable, including hires passes.

• Older GPUsUpdate

  GTX 10 and 16 series and RTX 20 cards lack bf16. Before February 2026 ComfyUI ran Anima in fp32 on them, very slowly or with a CUBLAS error. Current ComfyUI uses fp16 there.

• MacUntested

  No Apple Silicon reports yet. The weights are bf16, not fp8, so they should load on current PyTorch.

• 8 GB graphics cards
• 12 GB cards
• Laptop GPUs
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   Anima is native. Errors about embed_tokens size mismatches or 'conv_in.weight' almost always mean an old ComfyUI. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the three files

   A model, the Qwen3 0.6B encoder and the Qwen-Image VAE from the list above.

3. 3

   Put them in their folders

   Model in diffusion_models, encoder in text_encoders, VAE in vae. Restart ComfyUI.

4. 4

   Open the template

   Pick Anima Base v1: Text to Image in the template browser. Load Diffusion Model gets your Anima file, Load CLIP the Qwen3 encoder. Its type is stable_diffusion in the template; cosmos or qwen make no difference in current ComfyUI.

5. 5

   Match the sampler to the file

   The template is set for Base: 30 steps, CFG 4, er_sde, simple. For Turbo, change it to 8 to 12 steps and CFG 1.

6. 6

   Prompt with tags or sentences

   Danbooru tags, plain English or both. The template’s prompts show the quality tags; the section below has the rest.

Settings that work.

Base and Aesthetic

Steps

30 to 50

CFG

4 to 5

Sampler

er_sde

Scheduler

simple

Size

1024 × 1024512² to 1536²

Negative

yes

Text encoder

Load CLIP

Latent

EmptyLatentImage

Turbo

Steps

8 to 12

CFG

1

Sampler

euler or er_sde

Negative

noneignored at CFG 1

The card calls er_sde a reasonable default: neutral style, flat colours, sharp lines. euler is a bit more creative and suits Turbo and Aesthetic, euler_a gives softer lines. For a painterly look the author likes the beta57 scheduler from the RES4LYF pack.

Prompting.

Lowercase tags with spaces, not underscores (score tags keep theirs). Order: quality and safety, then 1girl or 1boy, character, series, artist, the rest. Put @ before artist names, or they barely register. Weights need to be higher than on SDXL, for example (chibi:2). For plain English, write at least two sentences and name each character before describing them.

masterpiece, best quality, score_7, safe, 

worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration

Leave the score_ tags out with Aesthetic. The safety tags are safe, sensitive, nsfw and explicit. Putting ye-pop or deviantart on the first line switches to the non-anime art styles.

How fast.

bf16 unless noted. The GTX 970 ran with fp16 compute. Turbo at 8 steps and CFG 1 needs a fraction of the model passes that 30 steps at CFG 4 do, since CFG above 1 runs two passes per step.

When it goes wrong.

size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151936, 1024])

ComfyUI is too old to know the Qwen3 0.6B encoder. Update it; the files are fine.

'conv_in.weight' when loading the model

Also an old ComfyUI. Update.

RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED on an older card

A GPU without bf16. Update ComfyUI (fixed in February 2026), or start it with --fp16-unet.

Hours per image on a GTX 1060 or similar

Same cause: fp32 fallback on a card without bf16. Update ComfyUI.

TAESD preview fails with a size mismatch

Use taew2_1 for previews.

lora key not loaded: lora_te_layers_…

A kohya LoRA with text encoder layers. Fixed in ComfyUI; update.

Plain, generic-looking pictures from Base

Base has a neutral default style by design. Add @artist tags, or use Aesthetic or Turbo.

Questions.

Sources: Anima model card, licence, ComfyUI Anima tutorial, Arc B580 report [1], SDXL comparison [2], GTX 970 and 1060 thread [3], ComfyUI issue #12477, sampler and hires thread.