Anima Receives Turbo v1.0 and Aesthetic v1.0

news

Anima, the 2B anime text-to-image model by CircleStone Labs, gets two new official variants: a fast Turbo version and a quality-tuned Aesthetic version.

Anima has received two new official variants — Turbo v1.0 and Aesthetic v1.0 — joining the existing Anima-Base v1.0 to offer three options for anime and illustration generation.

What is Anima?

Anima is a 2 billion parameter text-to-image diffusion model by CircleStone Labs, fine-tuned from NVIDIA's Cosmos-Predict2-2B-Text2Image. It is trained on millions of anime images and approximately 800,000 non-anime artistic images, with a knowledge cut-off of September 2025. No synthetic data was used in training.

The model uses a Qwen 3 0.6B text encoder and a Qwen Image VAE, both available as separate weight files. It supports resolutions between 512×512 and 1536×1536 pixels.

The Three Versions

Anima-Base v1.0

The pretrained, unrefined base model offering maximum flexibility, diversity, and style adherence. This is the recommended version for training LoRAs, as it has no aggressive aesthetic tuning to overcome.

  • Steps: 30–50
  • CFG: 4–5
  • Recommended samplers: er_sde, euler_a, dpmpp_2m_sde_gpu, euler

Anima-Turbo v1.0 (New)

A distilled version designed for fast generation. The distillation process also increases stability and gives the model a strong default style, though with reduced diversity.

  • Steps: 8–12
  • CFG: 1
  • Best for: Quick prompt iteration, cost-sensitive online platforms
  • On average, only slightly worse than Aesthetic while being dramatically faster

Anima-Aesthetic v1.0 (New)

Fine-tuned on high-quality images only, with quality tags stripped from training captions. This version delivers better consistency and a higher quality default art style. A v1.0b variant is also available — an aesthetics-only fine-tune without additional style LoRAs blended in.

  • Steps: 30–50
  • CFG: 4–5
  • Best for: Quality-focused generation where speed is less critical

Prompting and Workflow

The model is natively supported in ComfyUI. Drag the workflow image below into ComfyUI to get started:

Anima ComfyUI Workflow

Model files go into their respective directories:

  • anima-*-v1.0.safetensors → ComfyUI/models/diffusion_models/
  • qwen_3_06b_base.safetensors → ComfyUI/models/text_encoders/
  • qwen_image_vae.safetensors → ComfyUI/models/vae/

Anima uses Danbooru-style tags and natural language captions. Recommended quality prefix: masterpiece, best quality, score_7, safe. For the Aesthetic variant, quality tags are optional since it was fine-tuned exclusively on high-quality data.

A comparison workflow (anima_comparison.json) is included with the model for evaluating outputs across different model architectures.

Availability

The model and all variants are available on Hugging Face:

Anima Receives Turbo v1.0 and Aesthetic v1.0 | ComfyUI Wiki