Create Consistent Characters from One Reference Image with Anima Reference Adapter

What Is Anima Reference Adapter?

Anima Reference Adapter is a set of custom ComfyUI nodes for generating an Anima image from one reference image and a scene-only prompt. It attempts to preserve who is depicted—including hair, eyes, clothing, accessories, and some drawing style—while allowing the prompt to control the pose, background, action, framing, and composition.

You do not need to include a character name, artist tag, or detailed appearance tags. This makes the adapter useful for original characters as well as known characters. The current version is a pilot release trained on 300 characters for 2,000 steps.

This example illustrates the intended workflow: visual identity comes from the reference, while a prompt such as full body, walking, hands in pockets, city, street, night, rain defines the new scene.

Key Features

• Single-image conditioning: Supply one image to transfer character identity and some stylistic qualities.
• Scene-only prompting: Describe the desired action, pose, location, framing, and lighting instead of repeating the character's appearance.
• Independent guidance: Adjust text CFG and reference guidance separately.
• APG and rescale correction: Control excessive saturation at stronger reference-guidance values.
• ComfyUI integration: The project provides a loader, reference-latent node, and custom guider.
• Original-character support: The method does not depend on a recognized character name.

How the Adapter Works

The reference image is fitted to the requested output dimensions with white letterboxing and encoded by the Qwen-Image VAE. It becomes a second, clean latent frame at timestep zero beside the noisy target latent.

The reference frame's RoPE coordinates are shifted beyond the target image grid. This discourages pixel-aligned copying, allowing the requested pose and composition to differ from the reference. A rank-32 LoRA containing 45.9 million parameters teaches Anima to read this additional frame. The adapter LoRA is active only when the reference is present.

Separate text and reference guidance let you balance prompt adherence against identity and style transfer. Adaptive projected guidance, or APG, is available to help keep colors controlled when reference guidance is high.

Requirements

Before installing the custom nodes, prepare a working ComfyUI installation and the model components required by Anima:

• Anima base v1.0
• Qwen3-0.6B text encoder
• Qwen-Image VAE
• Anima Reference Adapter pilot weights

The Anima components are available from the CircleStone Labs Anima repository. The adapter weights are available from the Anima Reference Adapter repository on Hugging Face.

Installation

1. Install the custom nodes

Open a terminal, move into ComfyUI's custom_nodes directory, and clone the repository:

cd ComfyUI/custom_nodes
git clone https://github.com/ElectronGirl/ComfyUI-Anima-Ref-Adapter.git

2. Download the adapter weights

Download anima_ref_adapter_pilot_r32.safetensors and place it in the following directory:

ComfyUI/models/loras/anima_ref_adapter_pilot_r32.safetensors

3. Install or bypass Patch Sage Attention

Both included example workflows use the Patch Sage Attention node from ComfyUI-KJNodes. Install KJNodes if you use SageAttention. Otherwise, select that node in the workflow and bypass it with Ctrl+B.

4. Restart ComfyUI

Restart the ComfyUI server after installing the nodes and weights. Reload the browser interface with Ctrl+F5 so the new nodes are registered.

Generate Your First Image

1. Open the basic workflow

Load workflows/anima_ref_adapter.json from the cloned repository into ComfyUI. A second workflow, workflows/anima_ref_adapter_pag.json, includes Anima Safe PAG and is useful after the basic setup is working.

2. Load a reference image

Choose a clear image of the character whose identity and style you want to transfer. The Anima Ref Latent node fits the image into the selected output dimensions, adds white letterboxing where necessary, and VAE-encodes it. It also creates the empty target latent and a half-size reference encode for compatible multi-resolution workflows.

3. Write a scene-only prompt

Use quality terms, the character's gender, solo, and scene tags. Do not add a character name, artist name, or appearance description. A general prompt structure is:

masterpiece, best quality, score_7, newest, highres, absurdres, safe, <gender>, solo, <scene>

For example, to place the referenced character on a beach at sunset, use:

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, cowboy shot, looking back, smile, beach, ocean, sunset

Other scene-only prompt ideas include:

• upper body, sitting, holding cup, drinking, cafe, indoors, window, sunlight
• portrait, from side, looking afar, cherry blossoms, petals
• full body, standing, snow, forest, winter
• upper body, reading, holding book, library, bookshelf

4. Apply recommended settings

For the base model, begin with these values:

• Compute dtype: bf16 through ModelComputeDtype
• Sampler: Euler
• Steps: 28 to 30
• Scheduler: simple
• Text CFG: 4.5
• Reference guidance: 4.5 to 6
• Correction: apg, or rescale with phi 0.5
• Reference offset: Keep ref_offset_h at its training value of 1.0

Use bf16 rather than fp16. According to the project documentation, Anima overflows in fp16.

5. Queue the workflow

Verify that the model, VAE, text encoder, reference image, adapter weights, and prompt are connected correctly. Queue the prompt to generate the target image.

Understanding the Custom Nodes

Anima Ref Adapter Loader

This node attaches the adapter LoRA to an Anima MODEL. Keep ref_offset_h at 1.0, which is the value used during training.

Anima Ref Latent

This node prepares the reference and target latents. It fits and letterboxes the reference, encodes it with the VAE, creates a half-size encode, and returns an empty latent at the selected output size.

Anima Ref Guider

This node produces the GUIDER used by SamplerCustomAdvanced. It provides separate controls for text cfg and ref_guidance, along with APG or rescale correction. It also runs post-CFG patches such as PAG.

Advanced Tuning Tips

Balance identity and prompt adherence

Reference guidance controls how strongly the result follows the reference. The value 1 provides weak, less expensive conditioning. Values from 4.5 to 6 generally retain more character detail and style, but increasing the value can also produce stronger saturation. APG is the recommended correction when using higher guidance.

Use optional LoRAs in the correct order

Ordinary LoRAs should be connected before the Anima Ref Adapter Loader. The gallery examples used an Anima context-detailer LoRA at approximately 0.3 and a base-slider LoRA at approximately 0.6. These additions are optional rather than installation requirements.

Add Anima Safe PAG

For a more advanced graph, open workflows/anima_ref_adapter_pag.json. The gallery configuration used Anima Safe PAG on block 18, with scale 1.3 over the interval 0.2 to 0.9.

Use the Anima turbo LoRA carefully

The Anima turbo LoRA also works with the adapter. Load it before the adapter loader, then use the turbo LoRA's normal step and CFG settings rather than the base model recommendations. The reference adapter itself was trained on Anima base v1.0.

Prefer single-resolution sampling

Euler with 28 to 30 steps and the simple scheduler is the recommended starting point. The README reports that SamplerSPEED's half-resolution stage is unreliable because the pilot adapter was not trained for that progressive sampling setup.

Improve the reference input

Use a reference where the important identity features are visible, but expect the generated pose and composition to follow the prompt rather than the reference. If an accessory or fine pattern disappears, increasing reference guidance may help, although the pilot model has known limits on small details.

Current Limitations

• Character identity transfers more reliably than drawing style.
• Strongly painterly styles may transfer poorly.
• Small patterns, scars, earrings, and unusual skin tones are often lost.
• Large garments may change color.
• The referenced character can occasionally reappear as an object within the scene.
• The adapter was trained only on images generated by Anima and inherits the base model's biases.
• Separate character and style references are not available in this pilot release.

The roadmap includes separate character and style images, a larger dataset of 2,000 new characters in three artist styles, better painterly-style and fine-detail transfer, and half-resolution training pairs for progressive samplers.

Licensing

The ComfyUI node code is distributed under the Apache 2.0 license. The adapter weights are a derivative of the CircleStone Model and use the CircleStone Non-Commercial License. Review the weight license before using generated work or integrating the adapter into a project. The repository states that the project is not affiliated with or endorsed by CircleStone Labs.

Conclusion

Anima Reference Adapter provides a practical way to move a character into new scenes without writing identity, artist, or appearance tags. Start with the basic workflow, use bf16, Euler sampling, CFG 4.5, and reference guidance between 4.5 and 6. Once the core workflow works, add APG, Anima Safe PAG, or optional LoRAs while keeping ordinary LoRAs ahead of the adapter loader.