Click the clothes in your picture, load a picture of the clothes you want, and Krea 2 repaints that area to match them. The rest of your picture stays as it was.
How to use it
Load your image. Optional: drag the blue handles to crop it first.
On Mask by Touch, click the clothes you want changed. They light up before you click. A second click adds a piece.
Load a picture of the clothes you want. In What to take from it, name the piece, like "white tank top", "dress" or "outfit".
Run. The prompt Krea 2 wrote shows in Show Text, above the result.
What you need
AusBoss nodes 2.11.2 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.
The Krea 2 Turbo models and the AnyPaint LoRA:
krea2_turbo_fp8_scaled.safetensors →
models/diffusion_models· 13.1 GBqwen3vl_4b_fp8_scaled.safetensors →
models/text_encoders· 5.24 GBqwen_image_vae.safetensors →
models/vae· 254 MBkrea2_anypaint_rank32.safetensors →
models/loras· 229 MB
SAM 3, which finds the clothes you click: sam3.1_multiplex_fp16.safetensors →
models/checkpoints· 1.75 GB
Tips
Pick all of the old piece. Both sleeves and the hem: a second click adds a piece. If the edge falls short, raise Grow under the picture: a little extra is better than too little. The new piece only fills the picked area, and whatever you leave out stays.
Check the name under the picture before you run. It says what was picked. If a click took the whole person instead of a piece, press Undo and click again.
Wrong thing picked? The arrows under the picture swap it for the next thing it could mean. Paint adds an area by hand, and Erase takes some away. Before AusBoss nodes 2.12.0 this was one button, Draw.
An open coat picked alone stays open. In my test an open fur coat swapped for a sweater came out as an open cardigan over her own top. Pick the top under it as well.
Leave Auto prompt on. The text encoder turns your words into a full prompt, the way v1.1 always did. In my test on three pairs the new piece came through 3 times with it on and once with it off.
Use a clothing picture with the piece in plain view. Someone else wearing it is fine.
Use the whole clothing picture. A tight crop on just the fabric can lose the piece. Keep the person wearing it in view.
Name only the piece you want. Other clothes in the clothing picture can mix in. In one test, the sequins from the shorts ended up on the top.
Straps and necklaces that touch the picked area get repainted too, even the parts you did not pick. They can disappear or turn into the ones in the clothing picture.
Colors follow the light in your picture, so a piece can come out a little lighter than in the clothing picture.
It doesn't always take. In my tests a plain black t-shirt came out as a white tank top on two seeds.
If the picked area touches the right edge of your picture, a thin light stripe can show on that edge. Crop a few pixels off that side.
The four boxes in the middle do the work. Double-click one to see inside.
About 15 seconds a run on an RTX 5090. A new picture needs about 20 seconds before things light up.
Want to say what goes there in words instead? Use Krea 2 Inpaint Masked.
Credits
Krea 2 is by Krea, under the Krea 2 Community License: commercial use is allowed if you make under US$1M a year (details). The AnyPaint LoRA is by yijunwang2, under the same license.
v1.2 (2026-10-14): click the clothes instead of painting over them, with Mask by Touch. Gray parts in the new clothes are fixed. New Auto prompt switch, and you can crop both pictures on their loaders. One more download, the SAM 3 file. Needs AusBoss nodes 2.11.0 or newer.
v1.1 (2026-10-08): notes with arrows show where to load your two pictures and where to type, and the prompt Krea 2 wrote shows on the top level, above the result. Lowering the sampler’s denoise no longer leaves a flat gray patch. It paints the same pictures as v1.0.
v1.0 (2026-10-01): first release.
Questions or bugs: GitHub issues or the comments here. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.
Description
You click the clothes instead of painting over them. They light up under your pointer, one click picks a piece, and a second click adds another. The arrows under the picture swap a pick for the next thing it could mean. Paint adds an area by hand and Erase takes some away (one Draw button before AusBoss nodes 2.12.0).
New Auto prompt switch. On, it works like before: the text encoder turns your words into a full prompt. Leave it on. Off, Krea 2 gets your words as typed, and the new piece often does not come through.
Fixed: gray parts in the new clothes. Krea 2 is now shown the picked area in the picture's own colour, the way its inpaint LoRA was trained, and copies the clothing picture's colours more closely.
You can crop both pictures right where you load them.
One more download: the SAM 3 file. Needs AusBoss nodes 2.11.0 or newer. Update first.
FAQ
Comments (13)
This is crazy, ABSOLUT AMAZING, thanks for the ultra good work, i love it
@AusBoss i only like to tell, for me is not a big problem but Sometimes I run into issues with pair_to_latent, where it gets stuck for a long while. If I then go into the node, it immediately continues.
thanks for telling me. how big is the picture you load, and which gpu do you have? and does the ComfyUI console say anything while it sits there? i'll try to make it happen on my end.
at the moment i use pictures 1920x1080, my GPU is 4070 TI, console i didn't check at the moment, i will do next time. What surprises me is that as soon as I click into the node, it immediately jumps forward and works perfectly.
thanks, that helps. i tested it here and opening the box doesn't change anything on the comfyui side: the run keeps the same pace before and after. so most likely the run isn't stuck, the page just doesn't redraw that box's progress until you click into it. if it ever really is slow at that step, look for "Ran out of memory when regular VAE encoding, retrying with tiled VAE encoding" in the console. that means your 12 gb card was full right then and comfyui switched to a slower encode.
In the process of setting up/haven't tried it BUT
wow... your set up guide is clear and easy to follow. Very rare! and encouraging!
love uuuu
I am having a hard time figuring out a way to increase the resolution of the inpainted area, bc im using 2-4k images and the results are blurry, what should I do?
the painted part is made at about 1 megapixel and then scaled up to fit your picture, so on 2-4k pictures it comes out soft. you can raise Target (double-click the Cut out the area box, then on Crop For Inpaint open Target size, extend) to 1.5 or 2, but in my tests that sometimes repainted the clothes wrong. i'm testing a better fix for big pictures now.
Are you able to create something that can take clothing/outfit from an animated reference (such as anime/3d art) and then convert that outfit into a real life outfit reference (just clothing against a white background) so that it can then be used as a reference for another edit?
i haven't built that yet, but it's a good idea: turn the outfit into a photo of just the real clothes on a white background, then use that as the reference here. i'll test it and post here if it works.
@AusBoss Yeah, I think is a nice idea too
update: my Cut Out + Remove workflow can get close. the version i'm testing has a lora that redraws the whole outfit on its own on a plain background, but it's not perfect yet (12 of my 16 test pictures came out right) and the live version doesn't have it yet. you might not need that middle step though: Turbo Multi-Image Edit and this one both take clothes from a second picture and put them on your person. i haven't tried an anime picture as that second picture yet, so let me know if you do.