AnticPan/Anima-Role-Ref-KV

Anima Role Ref KV

v1.0 · Reference-guided anime illustration for Anima Base v1.0

中文

Generate new poses, outfits, expressions, and scenes from one character reference and a text prompt. No per-character training or trigger word is required.

The model encodes the reference image into a KV cache once per generation and reuses it across denoising steps to guide character appearance.

NSFW content was filtered out of the training data and was not used to train this release.

ComfyUI

1. Put the ComfyUI weights in ComfyUI/models/loras/.

2. Copy the comfyui folder to ComfyUI/custom_nodes/anima-role-ref-kv/ and restart.

3. Install the base components from Anima:

   File	Folder under ComfyUI
   anima-base-v1.0.safetensors	models/diffusion_models/
   qwen_3_06b_base.safetensors	models/text_encoders/
   qwen_image_vae.safetensors	models/vae/

4. Drag in the workflow, upload a reference, edit the prompt, and Queue.

The Anima Role Ref KV node loads the adapter itself; connect the original base model directly. An API workflow is also available.

Diffusers

Use Python 3.12. Download this repository and Anima Base v1.0 Diffusers, then run:

pip install -r requirements-diffusers.txt
python infer.py \
  --base-model /path/to/Anima-Base-v1.0-Diffusers \
  --reference reference.png \
  --prompt "masterpiece, best quality, 1girl, solo, upper body, smile, garden" \
  --seed 0 --output result.png

The entry point uses the Diffusers weights by default. Both formats contain the same adapter; reference generation requires the included node or Python runtime. Base components are downloaded separately.

Python import: from anima_role_ref_kv import AnimaRoleRefKVPipeline.

Settings

Start with 28 steps, CFG 4, BF16, and roughly 1 megapixel: 1024 × 1024, 832 × 1216, or 1216 × 832. Multiple aspect ratios are supported; dimensions must be divisible by 16. The workflow defaults to 832 × 1216.

Use a clear, unobstructed single-character reference and describe the desired pose, outfit, and scene. Details may drift with difficult references or strong prompt changes. Other base models, quantized weights, and additional adapters have not been validated.

License

An independent modified derivative of Anima, subject to the CircleStone model license; not an official CircleStone Labs product. Rights to the underlying model are granted directly by CircleStone Labs. Code licenses and attribution are in NOTICE.txt.