ElectronGirl/anima-ref-adapter

Anima Reference Adapter (pilot, r32)

Character and style from one image. Pose, place and framing from your prompt.

A LoRA + ComfyUI recipe for Anima: give one reference image and a prompt that only describes the scene. Hair, eyes, outfit and accessories come from the image; pose, background and composition follow the prompt. No character name, artist or appearance tags needed.

This is not a plain LoRA. The reference must be attached as an extra latent frame with shifted RoPE. Use the nodes and workflows from ElectronGirl/ComfyUI-Anima-Ref-Adapter.

🔜 Coming soon: separate character and style references. Take the character from one image and the drawing style from another, or either one alone. See the roadmap.

Gallery

Prompt: masterpiece, best quality, score_7, newest, highres, absurdres, safe, <gender>, solo, <scene>; Euler 28 steps, cfg 4.5, reference guidance 4.5 + APG, context detailer LoRA 0.3, base slider LoRA 0.6, Anima Safe PAG (block 18, scale 1.3, 0.2-0.9). Not cherry-picked.

Reference A input image	full body, walking, hands in pockets, city, street, night, rain	upper body, sitting, holding cup, drinking, cafe, indoors, window, sunlight	cowboy shot, looking at viewer, smile, beach, ocean, sunset
	portrait, from side, looking afar, cherry blossoms, petals	upper body, reading, holding book, library, bookshelf	full body, standing, snow, forest, winter

Reference B input image	full body, walking, smile, city, street, night, neon lights	upper body, sitting, eating, cake, cafe, indoors	cowboy shot, looking back, smile, beach, ocean, sunset
	portrait, closed eyes, smile, cherry blossoms, petals, wind	full body, sitting, on chair, crossed legs, classroom, sunlight	upper body, holding umbrella, rain, street, looking at viewer

Held-out characters

These characters were excluded from training. Each row: the reference image, the adapter's result for a different scene, the same scene-only prompt without the adapter (a generic character), and what the base model draws when the prompt also names the character with its appearance tags.

Scene prompts used in this figure

character	scene tags
Tohru	full body, looking at viewer, heart hands, crying, simple background, white background
Eternity Larva	full body, looking at viewer, against wall, closed eyes, gradient background, sunlight
Meiko	portrait, from side, writing, crying, beach, ocean, light rays
Crocodile	upper body, from below, drinking, sad, bookshelf, indoors

Quick start (ComfyUI)

1. Put anima_ref_adapter_pilot_r32.safetensors in ComfyUI/models/loras/.
2. git clone https://github.com/ElectronGirl/ComfyUI-Anima-Ref-Adapter into ComfyUI/custom_nodes/, restart.
3. The example workflows use Patch Sage Attention from ComfyUI-KJNodes; install it or bypass that node.
4. Open workflows/anima_ref_adapter.json, load a reference image, write a scene-only prompt.

	recommended	notes
base model	Anima base v1.0	no turbo LoRA (the adapter was trained on the base model)
compute	bf16 (ModelComputeDtype)	Anima overflows in fp16
sampler	Euler, 28-30 steps, simple	single-resolution samplers; SamplerSPEED's half-resolution stage is unreliable
cfg	4.5	text guidance
ref_guidance	4.5 - 6	1 = weak and cheaper; higher = more character detail and style, more saturation
correction	apg	or rescale (phi 0.5)
prompt	quality + gender + solo + scene tags	no character name, artist tag or appearance tags
extras (optional)	Anima context-detailer LoRA ~0.3, base-slider LoRA ~0.6, Anima Safe PAG	used for the gallery; ordinary LoRAs go in front of the adapter loader

How it works

The reference image is letterboxed to the output size, VAE-encoded and appended as a second frame on the DiT's time axis. It is fed clean (timestep 0) and its RoPE h coordinates are shifted by one target height, so the video prior cannot copy it pixel-aligned (no pose copying). A rank-32 LoRA (45.9M parameters, all attention and MLP projections of the 28 blocks), applied only to calls that carry the reference, learns to read it. Guidance separates text and reference scales, with adaptive projected guidance against over-saturation. Full details: report.md.

training
data	3,600 images generated by Anima: 600 groups of one character in one artist style x 6 scenes; 300 characters, 427 artists; 12 characters held out
pairs	17,280 (reference, target) pairs; target prompt keeps scene tags only (70%), + some appearance tags (20%), or nothing (10%)
optimization	flow matching, 2,000 steps, batch 4, AdamW 1e-4 cosine, bf16; 2.4 h on one RTX PRO 6000 Max-Q

Limitations

• Pilot release: 300 training characters, 2,000 steps, rank 32.
• Identity transfers much better than style; strongly painterly styles barely transfer.
• Small patterns, scars, earrings and unusual skin tones are often lost; large garments can drift in color.
• The reference character occasionally reappears as an object in the scene.
• Trained only on images generated by Anima; it inherits the base model's biases.

Artist names were used only as generation tags. Please respect artists: do not pass off imitations as their work, and follow the use restrictions in §4 of LICENSE.md.

Roadmap

	status
Single reference: character + style from one image	available (this release)
Separate character and style references: character from one image, drawing style from another, or either alone	coming soon
Larger training set: 2,000 new characters, each in 3 artist styles	generating
Better transfer of painterly styles and fine details (low-noise training, more data)	planned
Half-resolution training pairs for progressive samplers (SamplerSPEED)	planned

License and attribution

Non-commercial. These weights are a modified derivative of the CircleStone Model (Anima base v1.0), distributed under the CircleStone Non-Commercial License (LICENSE.md, NOTICE.txt). This is an independent project, not an official product of CircleStone Labs and not endorsed by it.

The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

Files

file
anima_ref_adapter_pilot_r32.safetensors	LoRA, bf16, ComfyUI key format (diffusion_model.*.lora_down / lora_up / alpha, 840 tensors)
report.md, figures/	technical report
workflows/	ComfyUI workflows
showcase/	gallery images
LICENSE.md, NOTICE.txt	CircleStone Non-Commercial License and attribution notice