Anima: 2B Parameter Anime Text-to-Image Model by Circlestone Labs
Anima is a 2 billion parameter text-to-image model by Circlestone Labs and Comfy Org, focused on anime concepts and artistic styles. Available in Base, Aesthetic, and Turbo variants for ComfyUI.
Anima by Circlestone Labs
Text-to-ImageAnimeArtisticCosmos-Predict2Anima is a 2 billion parameter text-to-image diffusion model created through a collaboration between Circlestone Labs and Comfy Org. It focuses on anime concepts, characters, and artistic styles, but can also handle a wide variety of non-photorealistic content. The model is trained on millions of anime images and about 800K non-anime artistic images, with no synthetic data used for training. The knowledge cut-off date for the anime training data is September 2025.
Versions
| Version | Description | Recommendation |
|---|---|---|
| Anima Base v1.0 | Pretrained, unrefined base model. Maximum flexibility, diversity, and style adherence. | Train LoRAs using this version |
| Anima Aesthetic v1.0 | Fine-tuned for better consistency and higher quality default art style. | Best overall quality |
| Anima Aesthetic v1.0b | Alternate aesthetic version — pure aesthetics full fine-tune without additional style adjustment and stabilization LoRAs merged in (unlike v1.0). | Niche |
| Anima Aesthetic v1.1 | Updated aesthetic-tuned variant | Latest aesthetic |
| Anima Turbo v1.0 | Distilled version for fast generations. Use at CFG 1 and 8-12 steps. | Recommended start — fast, stable, slightly lower quality than Aesthetic |
| Anima Preview | Early preview version | Development |
| Anima Preview 2 | Updated preview version | Development |
| Anima Preview 3 Base | Latest preview built on base variant | Development |
Installation
The model is natively supported in ComfyUI. Place files in their respective folders:
| File | Destination |
|---|---|
anima-*.safetensors | ComfyUI/models/diffusion_models/ |
qwen_3_06b_base.safetensors | ComfyUI/models/text_encoders/ |
qwen_image_vae.safetensors | ComfyUI/models/vae/ |
Generation settings
- Resolution: 512² to 1536² pixels
- Steps: 30-50 (Base/Aesthetic), 8-12 (Turbo)
- CFG: 4-5 (Base/Aesthetic), 1 (Turbo)
- Recommended samplers:
er_sde(neutral default),euler_a(softer lines),dpmpp_2m_sde_gpu(more creative),euler(good with Turbo/Aesthetic) - Scheduler tip:
beta57(ComfyUI RES4LYF custom node pack) can improve textures for more realistic/painterly looks
Prompting
The model is trained on Danbooru-style tags, natural language captions, and combinations of both.
Format
[quality/meta/year/safety tags] [1girl/1boy/1other etc] [character] [series] [artist] [general tags]Quality tags
- Human score based:
masterpiece, best quality, good quality, normal quality, low quality, worst quality - PonyV7 aesthetic model based:
score_9, score_8, ..., score_1 - Use either system, both, or neither
Recommended presets
- Positive prefix:
masterpiece, best quality, score_7, safe, - Negative:
worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration
Tips
- Use lowercase for tags, spaces instead of underscores (score tags use underscores)
- When a tag differs between Danbooru and Gelbooru, prefer Gelbooru
- Prompt weighting works but needs higher values than SDXL:
(chibi:2) - Aesthetic version: Don't use score_* tags — it's already high quality, score tags push it too far into slop