Introduction

Just another amateur remix of the Anima model.

With this model, you can create many fascinating artworks, just like an amateur.

Required Components

1. Text Encoder: https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/text_encoders/qwen_3_06b_base.safetensors?download=true

1. VAE: https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/vae/qwen_image_vae.safetensors?download=true

Alternative VAE: QwenimageVAE_liquid1087 or Qwenimagevae_liquid3844

Recommended Settings

Sampler: Euler A / ER SDE / dpmpp_2m_sde_gpu / res_multistep

Scheduler: Simple / Beta(57)

Steps: 30-50 steps

CFG: 4-5

If you're using TURBO for ANIMA or Cosmos-Predict2.5-2B base distilled extracted DMD2 LoRA, you might start off with 12 steps with CFG 1.0.

Known Issues / Limitations

1. Unless explicitly specified, version 1.0 is likely to produce characters with red eyes, especially under dark environment.

2. Stacking too many style LoRAs on top of this model might lead to over-saturation.

3. Malformed limbs or fingers on all versions - still figuring out how to fix. Adding 'bad anatomy' to the negative prompt seems better.

4. No up-to-date character knowledge - All character knowledge was acquired from the merged models. No plans for further improvement on this aspect.

Tips for Using LoRA with 2.9b

Latest version of Forge Neo supports Anima-2.9b and will do LoRA mapping automatically. For ComfyUI users, you might need to install ComfyUI-Anima-2.9B-loraPatch or ComfyUI-Anima-3.8B-loraPatch for remapping LoRAs. Alternatively, you can use the LoRAs that is natively built for Anima-2.9b, for example, Anima-2.9B Turbo LoRA Remapped.

License

'Non-Commercial Use Only' under section 2(b):

https://huggingface.co/circlestone-labs/Anima/blob/main/LICENSE.md