CalamitousFelicitousness/Anima-Preview-2-sdnext-diffusers

Anima Preview V2 — Diffusers Conversion

Unofficial diffusers-format conversion of circlestone-labs/Anima Preview V2 for use with SD.Next.

What Is This

This repo contains the Preview V2 weights converted to the HuggingFace diffusers layout so they can be loaded directly with from_pretrained. The custom pipeline (AnimaTextToImagePipeline) and all non-transformer components (text encoder, VAE, tokenizers, LLM adapter, scheduler) are identical to the Preview V1 diffusers conversion — only the transformer (diffusion model) weights differ.

Preview V2 Improvements

From the upstream model card:

• Retraining with better hyperparameters — makes the model more robust to finetuning
• Extended medium-resolution training — longer training at medium resolutions for more character knowledge
• Regularization dataset — improves natural language comprehension and preserves non-anime knowledge
• Base model philosophy — no aesthetic tuning; maximum knowledge breadth over aesthetic consistency

Usage with SD.Next

Select this repo as your model in SD.Next — it will be loaded automatically via the Anima pipeline.

Generation Settings

• Resolution: ~1 MP (1024×1024, 896×1152, 1152×896, etc.)
• Steps: 30–50
• CFG: 4–5
• Samplers: er_sde (default), euler_a, dpmpp_2m_sde_gpu

Credits

• Model: CircleStone Labs & Comfy Org
• Architecture: Derivative of NVIDIA Cosmos-Predict2-2B with Qwen3-0.6B text encoder
• Conversion: Automated via diffusers.CosmosTransformer3DModel.from_single_file()

License

Non-commercial use only — see LICENSE.md. Subject to CircleStone Labs Non-Commercial License and NVIDIA Open Model License Agreement.