ComfyUI-Anima-NAG

ComfyUI-Anima-NAG

日本語版 README

Experimental Normalized Attention Guidance node for the official Anima base model in ComfyUI.

Common SD/SDXL NAG nodes patch BasicTransformerBlock.attn2, but Anima is detected by ComfyUI as image_model="anima" and uses the Anima/Cosmos Predict2 attention path. This custom node wraps ComfyUI's optimized_attention path instead, so it can affect Anima cross-attention.

Anima is a 2B text-to-image model from CircleStone Labs and Comfy Org, focused on anime, illustration, and other non-photorealistic artwork. The official release is natively supported by ComfyUI and is available from:

• Civitai: Anima base-v1.0
• Hugging Face: circlestone-labs/Anima

This node can be used with the official base model to add negative-guidance control through Anima's cross-attention path. It can also be used together with Anima Turbo LoRA-style low-CFG workflows when you want NAG control at CFG 1.

Model files

For the official Anima base release, place the files in ComfyUI's model folders as follows:

• anima-base-v1.0.safetensors -> ComfyUI/models/diffusion_models
• qwen_3_06b_base.safetensors -> ComfyUI/models/text_encoders
• qwen_image_vae.safetensors -> ComfyUI/models/vae

Node

Anima Normalized Attention Guidance

• Category: Anima/Guidance
• Input:
  • model (MODEL): Anima model
  • scale (FLOAT): NAG strength
  • tau (FLOAT): normalization threshold
  • alpha (FLOAT): blend between original positive attention and normalized guided attention
  • start_percent / end_percent (FLOAT): active sampling range; 0.0 is the first step and 1.0 is the last step
  • only_anima (BOOLEAN): apply only when the model is detected as image_model="anima"
  • optimize_outside_range (BOOLEAN): compute only the positive branch outside the active NAG range for faster CFG 1 turbo workflows
• Output:
  • model (MODEL): patched model

Usage

1. Load anima-base-v1.0.safetensors with ComfyUI's Anima workflow.
2. Load qwen_3_06b_base.safetensors as the text encoder and qwen_image_vae.safetensors as the VAE.
3. Connect the Anima model to Anima Normalized Attention Guidance.
4. Connect the patched model output to your sampler.
5. Connect both positive and negative conditioning to the sampler.

Example settings:

• Anima Turbo LoRA: enabled
• sampler CFG: 1.0
• sampler steps: 8-12
• scale: 2.0
• tau: 2.5
• alpha: 0.5
• start_percent: 0.0
• end_percent: 0.5
• only_anima: true
• optimize_outside_range: true

These settings assume the Anima Turbo LoRA recommended 8-12 sampler steps. If you use a different step count or a different acceleration method, review start_percent, end_percent, scale, tau, and alpha for that workflow.

Sample workflow

A sample workflow is included at:

• WorkFlow/AnimaTurboLoRAwithNAG.json

This workflow demonstrates using Anima Turbo LoRA together with Anima Normalized Attention Guidance at CFG 1. It is intended as a starting point for the 8-12 step turbo setup described above.

Prompting notes

The official Anima model accepts Danbooru-style tags, natural language captions, and mixtures of both.

• Recommended positive prefix: masterpiece, best quality, score_7, safe,
• Recommended negative: worst quality, low quality, score_1, score_2, score_3, artist name
• Use lowercase for tags and spaces instead of underscores, except score tags such as score_7.
• A typical tag order is quality/meta/year/safety tags, subject count, character, series, artist, then general tags.
• Artist tags should use an @ prefix.

Notes

• This is experimental and tuned for Anima's current ComfyUI implementation.
• start_percent / end_percent are converted internally to the model's sigma range with ComfyUI's percent_to_sigma().
• With optimize_outside_range=true, partial ranges can speed up because steps outside the NAG range use the same positive-only behavior as ComfyUI's CFG 1 optimization. This option is intended for CFG 1 turbo workflows.
• The node needs a CFG batch containing both positive and negative branches. If no negative conditioning is connected, it will have no effect.
• Masked attention calls are skipped to avoid shape-specific mask slicing issues.
• If another node already set optimized_attention_override, this node wraps the previous override rather than replacing it outright.

Install

Place this folder under ComfyUI/custom_nodes/ and restart ComfyUI.

cd path\to\ComfyUI\custom_nodes
git clone <your-repo-url> ComfyUI-Anima-NAG

License

MIT