guide.en.md

Anima Reference Adapter: user guide

English · 한국어 · 日本語

This guide walks through installing the adapter in ComfyUI, making a first image, writing prompts, using two characters, tuning the settings and fixing common problems. For the method and measurements, see the README.

1. What you need

• ComfyUI that runs Anima. If you can already make images with Anima, you have everything.
• The Anima files from circlestone-labs/Anima:
  • anima-base-v1.0.safetensors in ComfyUI/models/diffusion_models/
  • qwen_3_06b_base.safetensors in ComfyUI/models/text_encoders/
  • qwen_image_vae.safetensors in ComfyUI/models/vae/
• About 1 GB of free VRAM on top of what Anima needs (the reference's cached keys and values at 1 MP). It was tested on an 8 GB RTX 3060 Ti.

The adapter weights do not need a separate download: the node fetches them on first use.

2. Install the nodes

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

Restart ComfyUI. Without git, download the repository as a ZIP from GitHub (Code → Download ZIP) and unpack it into ComfyUI/custom_nodes/.

3. Your first image

1. Download anima_ref_adapter.json and drag it onto the ComfyUI canvas. It is the standard Anima text-to-image workflow with one extra node, Anima Ref Adapter, between the model loader and the KSampler.
2. In Load Image, choose your reference image.
3. In the positive prompt, keep the quality tags and describe only the scene (see section 4).
4. Queue. The first run downloads anima_ref_adapter_v2_r64.safetensors (183 MB) into ComfyUI/models/loras/; later runs start right away.

4. Writing the prompt

The reference supplies the character (face, hair, eyes, outfit, accessories) and much of the drawing style. The prompt supplies everything else.

write	do not write
quality tags: masterpiece, best quality, score_7, safe,	the character's name
1girl or 1boy, and solo	artist tags
framing: cowboy shot, full body, upper body, portrait	hair, eye and outfit tags (a few are fine)
pose and action: sitting, lying, on side, holding cup
expression: smile, sleepy, laughing
place, time and light: cafe, indoors, day, beach, sunset

Example: masterpiece, best quality, score_7, safe, 1girl, solo, cowboy shot, sitting, cafe, holding cup, smile, looking at viewer, indoors, window, day

Tips:

• Write a pose that differs from the reference's. The adapter is trained not to copy the reference's pose, and a clear pose in the prompt helps.
• Parts the reference does not show (the legs in a bust-up image, the back of an outfit) are guessed by the model. A reference that shows the whole character gives more consistent outfits.
• Anima's negative prompt from its model card works as is.

5. Two characters

1. Open anima_ref_adapter_two_characters.json. It has two Load Image nodes joined side by side by the core Image Stitch node.
2. Load the left character in the first image and the right character in the second.
3. Replace solo with 2girls (or 1boy, 1girl, 2boys) and say who is where, with a few words each: 2girls, the girl on the left is a policewoman with black hair over one eye, the girl on the right is a mecha girl with aqua hair and mechanical wings, standing side by side, peace sign, smile, city, street, day
4. Use a landscape size such as 1216x832.

The two drawing styles tend to merge into one; the characters themselves stay apart.

6. Settings

All on the Anima Ref Adapter node.

setting	default	what it does
ref_guidance	4.5	How hard to push toward the reference. Raise it (5-6) when the character drifts; lower it when colors get too strong. At 1 there is no extra model call, so it is faster but weaker.
correction	apg	Keeps colors in check at high guidance. rescale is an alternative; none turns it off.
ref_end	1.0	The point in the schedule after which ref_guidance stops (the reference itself stays). About 0.6 is faster with nearly the same image.
ref_megapixels	1.0	Size the reference is resized to. Lower it (0.5-0.7) to save VRAM.
strength	1.0	Strength of the adapter LoRA.

Other LoRAs (style, detail) go in front of the adapter node, between the model loader and Anima Ref Adapter.

7. Making it faster

method	how	effect (RTX 3060 Ti, 832x1216)
turbo LoRA	a turbo LoRA for Anima base in front of the adapter; 10 steps, cfg 1, ref_guidance 1	67 s → 11 s; the reference style weakens
ref_guidance 1	on the node	67 s → 50 s
SageAttention	KJNodes Patch Sage Attention in front of the adapter	about 7%
EasyCache	the core EasyCache node after the adapter	about 7%
ref_end 0.6	on the node	about 7%

8. Troubleshooting

problem	what to do
The result looks like an ordinary Anima image, not the reference	Check that the KSampler's model comes from Anima Ref Adapter, not from the model loader directly. Raise ref_guidance.
Colors are too strong or burnt	Lower ref_guidance (3-4) and keep correction on apg.
The pose copies the reference	Write a clear, different pose and framing in the prompt.
The reference character appears again as a poster or a figure in the scene	A known limitation; change the seed or lower ref_guidance.
The download fails	Download anima_ref_adapter_v2_r64.safetensors by hand into ComfyUI/models/loras/.
Out of memory	Lower ref_megapixels, or use a smaller output size.
You want the pilot adapter	Use anima_ref_adapter_advanced.json: the pilot needs the reference at the output size.

9. Forge Neo

An extension for Stable Diffusion WebUI Forge (Neo) is available: sd-forge-anima-ref-adapter.