AI Style Transfer From Your Own Art in ComfyUI: Kaloscope Ships 44,129 Artist Names, and the Reference Image Is the Feature Worth Using
Most image generators make you describe the look and the subject in the same sentence. "A lighthouse at dusk, loose gouache, muted teal palette, visible paper texture." The words fight. Push the style harder and the lighthouse turns into a smudge. Describe the lighthouse carefully and the gouache fades into generic digital paint.
A free ComfyUI add-on called Kaloscope splits those two jobs apart. The prompt describes what is in the picture. The style comes in through a separate control, with its own strength setting. You can feed that control three ways: type an artist's name from a built-in list, blend several names with weights, or hand it a single image and say "look like this."
The list is what gets attention. It holds 44,129 artist names. The model card says nothing about where its millions of training samples came from, and nothing about the artists on that list.
The third option, one reference image, gets less attention. It is the one I would build a workflow around, because it works with the only style you have every right to: your own.
What Kaloscope is
Kaloscope is a small style add-on, about 50 MB per file, released on Hugging Face by heathcliff01 on October 23, with ComfyUI nodes from the Chenkin-x lab. It does not make images by itself. It plugs into one of three image models you already need to have installed: Krea 2 Raw, Krea 2 Turbo (the fast version) or MOSS-Image-Base. Each model gets its own matching file, and the files cannot be swapped between models.
Two newer MOSS versions followed the first release: one on October 25 and another, called step17000, on October 26, which the card says tells artists apart better and fixes some dark, blurry results.
Under the hood, a small image-reading model looks at your reference picture and turns its look into a set of numbers. Those numbers steer the generator through their own channel while the prompt carries the subject. That is why you get two separate guidance settings, one for how closely the image follows your words and one for how closely it follows the style.
The licence map, in plain words
Kaloscope is several pieces under several licences, and the card and the code repository spell out which is which.
The MOSS add-on is Apache 2.0. That is a permissive licence that allows commercial use.
The Krea add-ons follow the Krea 2 Community License, including its own commercial terms and usage policy. Read it before using the Krea versions for paid work.
The image-reading model and the artist-name bank follow the DINOv3 License, Meta's licence for the vision model they are built on, which comes with attribution and use conditions.
The ComfyUI nodes are GPL-3.0, a licence that matters if you redistribute modified code, not if you just make pictures with it.
What no licence tells you is where the training data came from. The card lists roughly 3.6 to 4.3 million training samples across versions and around 44,000 artist labels, and publishes the names in a text file. It does not say whose permission covers those images. That gap does not show up as a legal clause. It shows up as a question a client might ask you, and you will not have an answer.
Why the reference image is the right route
Typing a famous illustrator's name into a style box is fast, and it puts you in the same place as every other person typing that name. Your work looks like a knock-off of someone else's, and if they find it, they will say so.
A reference image of your own work flips that. Feed it one finished piece, a page from your sketchbook, a painted background from your last job, and the generator borrows your palette, your edge quality and your texture for whatever subject you describe. You are not borrowing anyone's identity. You are using a machine to make more first drafts in a look you already own.
That is useful in very ordinary ways. Storyboard roughs that already match the final art direction. Background plates for an animatic. Twenty quick variations of a cover layout so you can pick a composition before you paint the real one. None of these replace your finished work. They replace the hours of throwaway drawing before it.
Put this into practice: your style as a separate dial
You need ComfyUI installed and updated to a version that supports MOSS-Image and Krea 2, plus one of the three base models from Comfy-Org. I would start with MOSS-Image-Base, because its Kaloscope file is the Apache-licensed one.
1. Install the nodes. Clone the Kaloscope ComfyUI node pack into your ComfyUI/custom_nodes folder and install its requirements with the same Python that runs ComfyUI.
2. Download the add-on. From a terminal:
hf download heathcliff01/Kaloscope-style-adapter --local-dir Kaloscope-style-adapter
python Kaloscope-style-adapter/install_models.py --comfyui "/path/to/ComfyUI" --variant moss --with-encoder
The --with-encoder part fetches the image-reading model, which only the reference route needs. The installer puts in the October 25 version by default. The newer step17000 file is in the repository but not installed for you: copy models/Kaloscope-MOSS-B8-D256-step17000.safetensors into your models folder yourself and select it in the Kaloscope Adapter Loader node.
3. Open the reference workflow. Restart ComfyUI and drag in workflows/moss-b8-step10000/Kaloscope_MOSS_B8_reference.json. Skip the files ending in _api; those are for scripts, not the editor.
4. Choose your reference carefully. One image, your own, with a clear look: a strong palette, a recognisable mark-making style. Load it in the Load Image node. The node pack describes the reference as style guidance rather than a starting picture to edit, so your prompt still decides what is in the frame. It is centre-cropped to a square by default, so keep the important texture in the middle.
5. Start small. The node pack suggests confirming the setup at 512 by 512 first, even though the sample workflows run at 1024. The card's starting settings for MOSS: prompt guidance 3, style guidance 6, 25 steps, add-on strength 1.0. Write the prompt as plain subject matter with no style words at all.
6. Find your strength. Fix the seed (the number that makes a generation repeatable), then run the same prompt at strength 1.0 and 0.8, as the card suggests. I would also run it once with style guidance at 0, which should give you something close to the plain base model for comparison. Lay the three side by side. That comparison tells you more in five minutes than any setting I could recommend.
7. Scale up only once it looks right. Move to your working size in steps that are multiples of 64.
If your graphics card runs short of memory, the node pack lets you set the style encoder loader's device to cpu, so the image reading happens on your processor and the card keeps its memory for generating.
Where it breaks
Faces, bodies and layouts can fall apart. Both the card and the node pack warn that strong style settings and unusual references can damage anatomy and composition. If hands and faces matter, keep strength lower and fix them by hand.
One reference image per style. The workflow takes one picture as the style source. If your style shows up differently in line work and in paint, pick the one closest to the job, or run them separately.
MOSS is the less sharp version. The card does not promise the MOSS add-on separates styles as cleanly as the Krea one. The October 26 version narrows the gap, by the authors' account.
The usual speed shortcuts are untested. Smaller compressed model files (FP8 and GGUF), ControlNet pose and depth guides, and other add-ons stacked on top are "not yet fully verified," per the card. The node pack also says not to stack extra guidance patches on its style node. Add one thing at a time.
Mixing parts breaks it. Each Kaloscope file only works with its own base model. A Krea file on MOSS, or a MOSS file on the MOSS-Image-SFT variant, will not match.
No hardware figures. Neither the card nor the node pack says how much graphics memory you need. Test on your own machine before you plan a deadline around it.
The training sources are unknown. Even on the reference route, the model learned what "style" means from data the card does not describe. Using your own reference keeps your output yours in look. It does not answer that question.
What this changes
Separating style from content sounds like a technical detail. In practice it is the difference between fighting your prompt and directing it. With Kaloscope, you can write "a lighthouse at dusk, a fishing boat in the foreground" and let a picture of your own work do the talking about how it should look.
The artist-name list will get the screenshots and the arguments. Let it. Pick one finished piece you are proud of this week, run the seven steps, and compare the 0.8 and 1.0 versions against your original. If the roughs feel like yours, you have a new first-draft tool. If they feel like someone else's, you have learned that before a client did.
- Sources: Kaloscope model card · Kaloscope ComfyUI node pack · MOSS-Image-Base · Krea 2 Raw