Anima Reference Adapter: Put Your Anime Character in Any Scene From One Drawing, No Training
A free add-on for the Anima illustration model reads one picture of a character and redraws them anywhere you describe. It copies the design well, the drawing style badly, and its license is more generous than its label.
Keeping one character looking like the same person across twenty pictures has been one of the most annoying problems in AI illustration for years. The usual fix is to train a small custom add-on on a dozen drawings of your character, which costs an afternoon, a capable graphics card, and a stack of drawings you may not have yet.
The Anima Reference Adapter, posted to Hugging Face on October 27 by a developer called ElectronGirl, skips all of that. You hand it one picture. You type a scene. The character walks into the scene wearing the same outfit, the same hair, the same eyes, and the pose and background follow your words instead of the reference.
That sounds like every "consistent character" pitch you have scrolled past. What makes this one worth your time is the paperwork. The author published a full training report alongside the file, including a figure titled "failures," and the license file says something the label does not. Read together, they tell you exactly what job this tool is good for.
What it does, in plain terms
Anima is a free anime and illustration model from CircleStone Labs and Comfy Org. It draws characters and stylised art well, and its own model card says it "will not work well at realism." The adapter is a small add-on file that sits on top of Anima and changes one thing: it lets the model look at a picture while it draws.
You give it a reference image of a character. You write a prompt that describes only the scene: where they are, what they are doing, the time of day. You do not describe the character's appearance, name them, or name an artist. The report shows the adapter carrying across hair, eye colour, outfits and accessories such as horns, a halo, wings and a maid headdress, while, in the author's words, "Pose, framing and background follow the prompt in all 18 evaluation cases."
That last part matters more than it sounds. Plenty of reference tools copy the picture too faithfully, so every output inherits the reference's pose and camera angle. This one was built specifically to avoid that.
Why it does not just trace your drawing
The mechanism is simple enough to explain without any maths.
The adapter feeds your reference to the model as a second, finished picture sitting beside the one being drawn. Then it deliberately shifts the reference's position markers off to one side, so the model cannot line it up pixel for pixel with the new image. The model can see what the character looks like. It cannot see where things were in the frame. So it borrows the design and invents the composition.
The training data explains the rest. The author generated 3,600 images of 300 well-known anime characters drawn in the styles of 427 artists, six pictures per character-and-style pairing, each with a different pose, background and lighting. The adapter learned from pairs: here is the character in one picture, now draw them in a different scene from a prompt that does not describe them. Twelve characters were held back from training entirely, so the tests show the adapter working on characters it had never seen.
Training took about 2.4 hours on one large workstation card. The point is not that you should train your own. The point is that this was a small, careful experiment, and the author is upfront that it is a pilot.
The catch that decides how you use it: design travels, style does not
Here is the finding I would build a whole workflow around.
The author tested the adapter on original characters by an artist the model had never seen. The character design transferred well: hairstyles, ornaments, outfits, colour palettes. The artist's painterly rendering did not. At low reference strength, the outputs slid back to what the report calls "a crisper default Anima look."
So treat this as a design-consistency tool, not a style-copying tool. If you are an illustrator with an original character, the adapter keeps their costume and silhouette steady across a sequence of scenes. It will not reproduce your brushwork. Your style has to come from the prompt, from Anima's own style handling, or from you painting over the result.
That is the better split for a working artist anyway. Your design is the thing you need repeated. Your hand is the thing you want to keep.
The license is better news than the label
The model card says the weights are non-commercial, and a lot of people will stop reading there. Read the license file instead.
The adapter is a derivative of Anima, so it carries the CircleStone Non-Commercial License. That license bars using the model file "for any commercial or production purposes" without a separate deal with CircleStone Labs. You cannot build it into a paid app, run it behind a paid service, or sell access to it.
The pictures are a different matter. Section 2(e) of the same license says CircleStone claims no ownership of what you generate and that you "may use Outputs for any purpose (including for commercial purposes)." Anima's own model card spells this out with examples: selling images, paid commissions, concept art for a paid game.
There is a condition attached. The license makes you solely responsible for using outputs in a lawful, non-infringing way. And this tool's whole purpose is to reproduce a character's design. Feed it your own original character and the pictures are yours to use. Feed it a famous character from a show, and the license has handed the legal risk straight to you.
I am not a lawyer, and if money is riding on it, have one read the license. But the practical rule is simple: the file stays personal, the pictures can go to work, and the reference should be a character you own.
Put this into practice: a character sheet in an afternoon
You need ComfyUI, the free node-based program most people use to run local image models, and a graphics card that already runs Anima. The adapter card does not list memory needs.
1. Get Anima working first. From the Anima model card, put the base model file (anima-base-v1.0) in your diffusion models folder, its text encoder in the text encoders folder, and its image decoder (the file that turns the model's work into visible pixels) in the vae folder. Use the base model, not Turbo. The adapter was trained on the base model and the card warns against pairing it with a speed add-on.
2. Install the adapter. Put anima_ref_adapter_pilot_r32.safetensors in ComfyUI/models/loras/. Clone the ComfyUI-Anima-Ref-Adapter node pack into custom_nodes/ and restart. This step is required: the reference enters as an extra picture, and an ordinary add-on loader will not pass it through. Install ComfyUI-KJNodes as well, or bypass the one node in the workflow that uses it.
3. Load the workflow. Open workflows/anima_ref_adapter.json from the adapter repo. The card's settings: the Euler sampler, 28 to 30 passes, a prompt-strength setting (CFG) of 4.5, and reference strength between 4.5 and 6 with the card's colour correction switched on (more on that in step 7). Keep the number-precision setting the card names, BF16; the card says Anima breaks in the other common half-precision mode, FP16.
4. Pick a clean reference. The report does not say which kind of reference works best, but the training references were single-character images on plain white padding. Start there: one character, full outfit visible, simple background.
5. Write scene-only prompts. Quality tags, gender, solo, then the scene: "on a rooftop at dusk, wind in hair," "reading in a library," "running through rain." Leave out hair colour, outfit and names. That is the adapter's job.
6. Run a dozen scenes and compare. Put every result next to the reference. Check the small things first: earrings, scars, patterns, unusual skin tones. Those are the details the report says drop out.
7. Watch the colour as you raise the reference strength. Higher reference strength pulls in more detail but pushes saturation. The report's example is a black hoodie drawn dark green. The report's fix is a colour correction setting called adaptive projected guidance (APG); the author says it allows reference strength 6 with less colour drift than strength 4.5 uncorrected, and recommends 4.5 to 6 with it switched on.
Where it breaks
Small details vanish. Patterns, scars, earrings, and pale or grey skin are often lost. If your character's identity lives in a tattoo, expect to paint it back.
The character sometimes becomes a prop. The report's failure figure shows the reference character turning up as an object in the scene, such as a plush toy. Reroll and move on.
Big garments drift in colour, especially at high reference strength.
Some samplers break it. The report says samplers that change resolution mid-render break the adapter, and shrinking the reference causes artefacts at high strength. Stick to one resolution.
It is a pilot. Three hundred characters, no quantitative scores yet, all training data generated by Anima itself, so it inherits Anima's habits and blind spots. The author plans an update that takes the character from one image and the style from another. Until then, style is your job.
No hosted demo. You need a local setup. There is no free browser version to try first.
The bigger point
Most character-consistency tools sell you on the idea that the problem is solved. This one tells you which half it solved. Design carries over from one drawing. Rendering style mostly does not. Colour gets louder as you push it. Tiny details fall off.
That honesty is the useful part, because it tells you where your own hands go. Let the adapter handle the boring consistency work, the same jacket in twenty scenes, and spend your time on the parts it cannot do.
If you have an original character sitting in a sketchbook, give it one afternoon. Draw one clean reference, write twelve scene prompts, and lay the results out as a sheet. You will know within an hour whether your design survives the trip, and the pictures that do are yours to use.