Anima Native Reference V2 — E180

Anima Native Reference V2 E180 is an experimental two-reference,
image-conditioned generation model developed as part of the
NOOB2 Project — Character Reference Bypass Injector Research.

It generates a new anime illustration from pure noise while conditioning on:

• a text prompt;

• Reference Image 1, assigned to ordered slot 0;

• Reference Image 2, assigned to ordered slot 1.

The two slots do not have permanently assigned scene, character, identity, or
style roles. Their order is preserved, while the prompt determines how the
model should interpret and combine them.

The Native Reference V2 modules are already integrated into the E180
checkpoint. No external LoRA, IP-Adapter, or separate reference-adapter weight
file is required.

This model uses bypass injection, preserving the original text-to-image capabilities of the Anima base model. You can continue using the standard ComfyUI nodes for regular text-to-image generation.

Quick Links

• Custom node and workflows:
  https://github.com/HK416-TYPED/ComfyUI-Anima-Native-Reference

• Hugging Face research repository:
  https://huggingface.co/LAXMAYDAY/NOOB2-Project-Character-Reference-Bypass-Injector-Research

• Anima text encoder and VAE:
  https://huggingface.co/circlestone-labs/Anima

What This Model Does

Anima Native Reference V2 accepts two ordered reference images and a prompt,
then creates a new image from pure noise.

Possible tasks include:

• using Image 1 as a scene reference and Image 2 as a character reference;

• combining character identity and costume information from both images;

• transferring a character into a new composition;

• preserving selected facial, hair, color, clothing, or accessory information;

• prompt-controlled changes to pose, framing, lighting, and environment;

• research into multi-reference character generation without training a
  separate LoRA for every character.

The role of each image is defined by your prompt rather than by a fixed node
assignment.

Important Architecture Notes

• This is not img2img.

• Sampling starts from pure noise.

• There is no denoise-strength control.

• It does not use the stock ComfyUI KSampler path.

• It does not expose standard MODEL / CLIP / VAE outputs.

• The E180 checkpoint already contains the learned Native Reference V2 modules.

• No external LoRA, IP-Adapter, or reference-adapter checkpoint is required.

• The current release requires exactly two reference images.

• Each reference input currently accepts one image only.

The two reference slots are:

Reference Image 1 → ordered slot 0
Reference Image 2 → ordered slot 1

Their semantics are determined by the prompt.

Installation

1. Install the Custom Node

⚠️ CUSTOM NODE REQUIRED

Download and install ComfyUI-Anima-Native-Reference before using this model:

https://github.com/HK416-TYPED/ComfyUI-Anima-Native-Reference

This is a complete integrated diffusion checkpoint, not a LoRA.
Do not load an additional stock Anima diffusion model alongside it.

Git installation

Open a terminal in ComfyUI/custom_nodes and run:

git clone https://github.com/HK416-TYPED/ComfyUI-Anima-Native-Reference.git
cd ComfyUI-Anima-Native-Reference
python -m pip install -r requirements.txt

Restart ComfyUI after installation.

Windows Portable ComfyUI

From the root of a standard Windows Portable ComfyUI installation:

cd ComfyUI\custom_nodes
git clone https://github.com/HK416-TYPED/ComfyUI-Anima-Native-Reference.git
cd ComfyUI-Anima-Native-Reference
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt

Restart ComfyUI after installation.

The custom node does not pin or replace PyTorch. Do not manually replace a
working CUDA installation of torch or torchvision.

2. Download the Required Model Files

You need three files.

Integrated Native Reference V2 checkpoint

Download the model file from this Civitai release:

anima-native-ref-v2-e180-step64080-256px.safetensors

Place it in:

ComfyUI/models/diffusion_models/

This is the complete diffusion-model checkpoint used for generation.

Do not also load a stock Anima diffusion checkpoint.

Anima Qwen3 text encoder

Download:

qwen_3_06b_base.safetensors

From:

https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders

Place it in:

ComfyUI/models/text_encoders/

Qwen-Image VAE

Download:

qwen_image_vae.safetensors

From:

https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae

Place it in:

ComfyUI/models/vae/

Your final model layout should be:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── anima-native-ref-v2-e180-step64080-256px.safetensors
    ├── text_encoders/
    │   └── qwen_3_06b_base.safetensors
    └── vae/
        └── qwen_image_vae.safetensors

3. Download the Workflow

The standard ComfyUI workflow and API workflow are available in the custom-node
repository:

• Standard workflow:
  https://github.com/HK416-TYPED/ComfyUI-Anima-Native-Reference/tree/main/example_workflows

• Local /prompt API workflow:
  https://github.com/HK416-TYPED/ComfyUI-Anima-Native-Reference/tree/main/api_workflows

The UI workflow and API workflow use different ComfyUI JSON formats and should
not be treated as interchangeable.

Included Custom Nodes

Anima Reference V2 Loader (E180)

Loads and caches:

• the integrated E180 Native Reference V2 checkpoint;

• the Anima Qwen3-0.6B text encoder;

• the Qwen-Image VAE.

The node returns a custom:

ANIMA_NATIVE_REF_V2_PIPELINE

Anima Reference V2 Generate (2 Refs)

Generates one image using:

• the Native Reference V2 pipeline;

• Reference Image 1;

• Reference Image 2;

• positive and negative prompts;

• seed;

• output width and height;

• steps;

• CFG;

• flow shift;

• reference scale;

• maximum preprocessing area per reference.

Basic Usage

1. Install the custom node.

2. Put the three required model files in the correct folders.

3. Restart ComfyUI.

4. Load the provided workflow.

5. Select one image for Reference Image 1.

6. Select one image for Reference Image 2.

7. Write a prompt explaining how both references should be used.

8. Queue the workflow.

A prompt can define Image 1 as a scene and Image 2 as a character, but this is
only one possible assignment. The slots do not have permanent roles.

Prompt Examples

Scene from Image 1, character from Image 2

Use the first image as the scene template and the second image as the
character template. Generate an edited result where the character matches
Image 2 and the scene remains faithful to Image 1.

General two-reference generation

Create a new anime illustration using both reference images. Preserve the
relevant visual information from Reference Image 1 and Reference Image 2 while
following this instruction: show the referenced character in a coherent new
composition with clean anime rendering and detailed eyes.

Combine character information from both images

Create a new anime illustration that combines the character identity and
costume details shown across both reference images. Preserve the most
recognizable facial features, hairstyle, colors, and accessories while placing
the character in a new full-body composition.

Keep Image 1 identity and use Image 2 for costume

Use Reference Image 1 for the character's face, hairstyle, and overall
identity. Use Reference Image 2 only for the costume and accessories. Create a
new full-body anime illustration in a different pose and environment.

Be explicit about which image should control identity, costume, scene, pose,
style, or other visual information.

Recommended Settings

SettingRecommended valueResolution256 × 256Steps40CFG1.0Flow shift5.0Reference scale1.0Maximum area per reference65,536 pixelsOrdered slot IDs[0, 1]Attention modetorchOffload modebalanced

The checkpoint was trained and validated with 256-pixel-area buckets.

Larger dimensions that are multiples of 16 are accepted by the current
implementation, but they should be treated as experimental. Higher resolution
does not guarantee better character consistency.

Memory Modes

balanced

Keeps the DiT on the GPU and moves the text encoder and VAE to the GPU only
during their respective phases.

This is the recommended default.

high_vram

Keeps the DiT, text encoder, and VAE on the GPU for faster repeated generation.

text_encoder_cpu

Runs text encoding on the CPU as a slower fallback when GPU memory is limited.

The node maintains bounded caches for the runtime, prompt embeddings, and
preprocessed reference latents. Changing a seed or sampling setting does not
normally require reading the full checkpoint from disk again.

Current Limitations

This is an experimental research-preview release.

• Exactly two reference inputs are currently required.

• Each reference input accepts one image only.

• Reference slots are ordered but do not have automatically inferred roles.

• The model may over-copy pose, framing, clothing, or background from a
  reference.

• Fine facial features may drift between seeds.

• Eye color, hair ornaments, and small accessories may change or disappear.

• Clothing information from the two references may become mixed.

• Complex backgrounds may leak into the generated result.

• Extreme poses and camera angles may reduce identity consistency.

• Multi-character inputs may cause identity mixing.

• The trained and validated scale is based on 256-pixel-area buckets.

• Larger outputs remain experimental.

• Photorealistic identity transfer is outside the primary target domain.

• Exact reconstruction of either reference is not guaranteed.

• Arbitrary stock Anima or future Native Reference checkpoints are not
  automatically supported.

Try multiple seeds and write explicit prompts when evaluating a reference pair.

Verified Integration

The packaged runtime was validated with the exact E180 release files on an
NVIDIA RTX PRO 6000 Blackwell.

Validation covered:

• loading the integrated E180 checkpoint;

• loading the Anima text encoder and Qwen-Image VAE;

• two-reference pure-noise generation;

• ordered slots [0, 1];

• Native Reference V2 sequence and attention paths;

• prompt and reference-latent caching;

• repeated runtime reuse;

• standard ComfyUI workflow validation;

• local /prompt API execution;

• generated-image saving;

• Chinese and Japanese Unicode prompt preservation.

For the held-out integration case, the ComfyUI output was pixel-identical to
the accepted native command-line output when decoded to RGB pixels.

These tests verify implementation and workflow parity. They are not a complete
benchmark of character-reference quality.

Responsible Use

Only use reference images that you own or are authorized to use.

Do not use this model to:

• impersonate real people;

• misrepresent generated content as authentic evidence;

• harass or defame individuals;

• violate privacy or publicity rights;

• reproduce private or sensitive imagery without permission;

• infringe copyright or contractual restrictions;

• create illegal or exploitative content.

Users are responsible for applying suitable moderation, human review, and
access controls in downstream applications.

License

The integrated model checkpoint, required Anima assets, and custom-node source
code are distributed under separate terms.

• The E180 checkpoint remains subject to this model release's terms, the
  applicable upstream Anima terms, and any relevant third-party licenses.

• The Qwen3 text encoder and Qwen-Image VAE remain subject to their upstream
  terms.

• The ComfyUI custom node and included runtime source are covered by the
  licenses included in the GitHub repository.

• A software license applying to inference code does not override the license
  governing the model checkpoint.

Review the current Civitai version license, Hugging Face repository terms, and
upstream Anima licenses before redistribution, derivative-model publication,
public inference hosting, or commercial use.

Acknowledgements

This project was made possible through the contributions of many developers,
researchers, and testers.

• Yidhar and Nebulae contributed to model training and dataset
  development.

• spawner provided important inspiration and portions of the code through
  the anima-edit project.

• potato contributed to the project's early development.

• GHOSTLXH, 年糕特工队, 轻松, Free Will, and 韩小强
  contributed testing, feedback, and early validation.

We also thank CircleStone Labs and Comfy Org for the development and
release of Anima, and the wider ComfyUI developer community for the
inference framework and ecosystem.

Special thanks to Comfy.org for providing GPU sponsorship.