Anima Role Ref KV

v2.0 · Character-reference generation for Anima Base v1.0

中文 · Examples · Evaluation · v1.0 archive

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

BF16 weights only (v2.0 and v1.0). Update the included Python runtime or ComfyUI custom node before loading — older FP32-only loaders reject BF16 files. Both versions live side-by-side by filename.

What changed in v2.0

v2.0 improves character resemblance and markedly reduces unintended washed-out outputs on natural-language prompts, with slightly lower overall tag adherence. Automated metrics, single training seed — see Evaluation.

Files

ComfyUI

1. Put the desired weight file (anima-role-ref-kv-v2.0.safetensors or anima-role-ref-kv-v1.0.safetensors) in ComfyUI/models/loras/.
2. Replace ComfyUI/custom_nodes/anima-role-ref-kv/ with the included comfyui folder and restart ComfyUI.
3. Install the base components from Anima (anima-base-v1.0.safetensors, qwen_3_06b_base.safetensors, qwen_image_vae.safetensors).
4. Load workflows/anima-role-ref-kv-v2.0.json (or v1.0.json), upload a reference image, edit the prompt, and Queue. You can switch between v2.0 and v1.0 directly in the Anima Role Ref KV node's adapter dropdown.

Diffusers

Use Python 3.12 with Anima Base v1.0 Diffusers:

pip install -r requirements-diffusers.txt
python infer.py \
  --base-model /path/to/Anima-Base-v1.0-Diffusers \
  --adapter diffusers/anima-role-ref-kv-v2.0.safetensors \
  --reference reference.png \
  --prompt "highres, safe, 1girl, solo, upper body, sitting, bench, park, tree" \
  --seed 20261106 --output result.png

Pass --adapter diffusers/anima-role-ref-kv-v1.0.safetensors to run v1.0. Both require the included reference runtime and are not standalone ordinary LoRA files.

Settings and prompt format

• Sampling: 28 steps, CFG 4, BF16, around 1 MP (832 × 1216, 1024 × 1024, or 1216 × 832, multiples of 16). Use a clear single-character reference.
• Tag prompts (training format): start with highres, safe, 1girl, solo, ... (or 1boy), then desired framing, pose, expression, outfit, background, and medium/style tags. Do not describe the reference character's existing hair/eye traits unless you want to override them.
• Natural-language prompts (training format): start with safe, 1girl. followed by concise sentences for framing, action, outfit, and scene (e.g. safe, 1girl. Upper body shot from the waist up. She is sitting on a bench in a park with trees.).
• Negative prompt: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration.

License

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