CalamitousFelicitousness/Anima-1.0-Base-Diffusers

Anima 1.0 Base (SD.Next Diffusers Conversion)

Diffusers-format conversion of Anima 1.0 Base for use with SD.Next.

Anima is a 2 billion parameter text-to-image model created via a collaboration between CircleStone Labs and Comfy Org. It is focused on anime concepts, characters, and styles, and on non-photorealistic illustration in general; it is not intended for realism. The Base version is the pretrained, unrefined base model, with maximum flexibility, diversity, and style adherence; its default style is plain and neutral, especially without artist or quality tags. LoRAs should be trained on this version.

Original model: circlestone-labs/Anima (split_files/diffusion_models/anima-base-v1.0.safetensors)

Architecture

• Transformer: CosmosTransformer3DModel (2B params, 28 layers)
• Text Encoder: Qwen3-0.6B (replacing Cosmos T5-11B)
• LLM Adapter: Custom cross-attention adapter bridging Qwen3 to the transformer
• VAE: AutoencoderKLWan

Recommended Settings

• 30-50 steps, CFG 4-5
• Resolutions between 512x512 and 1536x1536

Prompting

• Trained on Danbooru-style tags, natural language captions, and combinations of both. Tags are lowercase with spaces instead of underscores; score tags are the only tags that use underscores.
• Recommended positive prefix: "masterpiece, best quality, score_7, safe, "
• Recommended negative: "worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration"
• Artist tags require an @ prefix (e.g. "@artist name"); without it the effect is very weak.

Finetuning

• The LLM adapter should not be trained (set llm_adapter_lr=0 or the trainer's equivalent); it strongly influences outputs and degrades easily.
• A low learning rate is recommended: around 2e-5 for a rank 32 LoRA, adjusted from there.

License

CircleStone Labs Non-Commercial License v1.2 (see LICENSE.md). As a derivative of Cosmos-Predict2-2B-Text2Image, the model is also subject to the NVIDIA Open Model License. The non-commercial restriction applies to the model weights, not to generated images.