Krea 2 Identity Edit
#v1.2 — new things
In the model:
Better likeness
Character sheets — use and creation
Head/face swap (+ eye/person replace) — thanks to stablellama
Outpainting
Inpainting
Try-on
Better person removal
Higher fidelity + a real high-res pass
In the nodes:
ref_boost— reference-fidelity dial ← don't miss this one, try between 2 and 6Fixed blurry images /
fitgeometry (also removes the old "match the aspect ratio" requirement)
**Settings:** Turbo, 8–12 steps (8 = composition, 12 = face detail, ~10 balanced), CFG 1.0, LoRA strength 1.0. Removals/deletions: Two-ref edits: **scene = image 1, person = image 2.**
Requires the [ComfyUI-Krea2Edit nodes](https://github.com/lbouaraba/comfyui-krea2edit) — see [CHANGELOG](https://github.com/lbouaraba/comfyui-krea2edit/blob/main/CHANGELOG.md).
Instruction-based, identity-preserving image editing for Krea 2 (12.9B single-stream MMDiT). Give it an image and a plain-language instruction; it edits while preserving what you didn't ask to change — including the person.
An unofficial community fine-tune of Krea 2 Raw. Not an official Krea product; not affiliated with or endorsed by Krea.ai, Inc.
Requires the ComfyUI-Krea2Edit node pack — the LoRA is trained with dual conditioning (in-context VAE tokens + image-grounded Qwen3-VL encoding) that stock nodes don't provide. Two ready-made workflows ship with it.
What it does
Person re-staging with likeness: "create a photo of this person at a night market" — same face, same outfit down to individual moles and marks, fully relit to the new scene. New camera angles and poses included.
Local edits: recolor, add/remove/replace objects, attribute and outfit changes, with near-pixel preservation of the rest of the frame.
Replace-with-reference: "replace the woman with a big orangutan" — the replace verb is trained, locality holds.
Full-image restyles: global style with preserved composition.
Two-input edits (experimental): scene + person as separate references. Outfits and placement work well; see limitations for faces.
Composes with your LoRAs: character/body/style LoRAs stack on top and steer the prior — something closed editors structurally can't offer.
Recommended settings
Task type Model Steps CFG Most edits (add, recolor, restyle, re-stage) Turbo 8 1.0 Removals / large deletions Raw 20 3.0
Match the output aspect ratio to the source image. Training pairs are same-size; AR mismatch degrades preservation (edits may apply to only part of the frame).
Generate at ≤2MP. Above that, source content can bleed or subjects duplicate (training was 768/1024-class).
grounding_pxis a real dial (trained range 512–1536): lower values = stronger edit adherence and more uniform scene changes; higher values = stronger identity/likeness. 768 is a balanced default; try 1024+ for people.At CFG > 1, ground the negative too (empty prompt + same image).
LoRA strength 1.0.
Known limitations (honest list)
Likeness is texture-faithful, proportion-conservative. Moles, skin character, hair, and lighting adapt beautifully; strongly distinctive facial geometry (unusual nose, eye spacing, face length) regresses toward typical proportions. People whose identity lives in texture and structure transfer best; geometry-defined faces read as a "close relative."
Two-person inputs keep outfits distinct but faces drift toward each other. Workaround that works today: chain single-ref inserts (place person A, then a second edit pass adding person B from their reference).
Removal works but is not yet reliable — always use the Raw/CFG 3 recipe; expect occasional re-renders instead of deletions.
Outfit swaps are hit-or-miss — changing what a person wears sometimes works cleanly and sometimes doesn't apply; reroll or rephrase.
Local edits aren't always perfectly local — add/remove/replace operations can sometimes alter other parts of the frame or shift the overall color grade. If preservation matters, compare against the source and reroll.
Highly unusual visual content (extravagant hairstyles, extreme body types) can drift toward the base prior — a subject LoRA stacked on top fixes this.
License
The LoRA weights are a Derivative Model of Krea 2 and are distributed under the Krea 2 Community License Agreement (see also NOTICE). Key points for users: commercial use is permitted under the license's revenue threshold (§2.3, currently <$1M/yr — above that, contact Krea for an enterprise license); deployments must implement reasonable content moderation (§4.2); AI disclosure obligations apply where required (§4.3). This repository modifies the Krea Model as permitted by §3; it is not endorsed by Krea.
Research/portfolio release by a self-funded hobbyist.
Showcase
All reference people below are themselves AI-generated — no real likenesses. Prompts are embedded in each image.
Description
FAQ
Comments (103)
omg i hope it's works well !
I hope you'll enjoy using this! If you do, I'd love to see your generations. known good generations with input and output image + prompt will help me to scale my training data and further improve this model. :)
@conrad_locke will
@conrad_locke ive done some tests and it works well. Like you said for the "clothes change" its a bit random, for the face and body consistency its not so bad ! Some times it duplicate the person in the scene.
I will upload some test this night. But very nice work !
Can i help you to build the v2 ?
@toto47677 If you have training data, past image edits paired with input image and prompts, then I'd love to include them in my next training run! V1.1 is currently training as we speak. I've restarted training from an earlier checkpoint with a slightly different setup in the hopes of increasing face likeness. If it works I'll potentially update this model later tonight.
The double character issue is known to me. And it most likely stems from me training at 512, but training at higher resolutions is currently not feasible for me sadly.
@conrad_locke which card do you have ? i have a 5090 and if needed 50 euro of credits or more on runpod for a train on a big gpu.
trained on raw means it should work on turbo too, no? your wf has raw
yes, I basically only use turbo. when removing very large objects or persons that take up a significant amount of the image I recommend using cfg3 and 20-30 steps. otherwise turbo is perfectly fine.
Oh my god, someone actually awnsers my prayers. You will be blessed.
Thank you! I hope I'll be able to improve this further :)
Suprisingly good results with Krea 2 turbo! Great job!
same. this is really good even with Turbo.
LOL Krea 2 is so good the community couldn't even wait for the official edit model.
Then again. After what happened with Z-Image and Qwen 2. Not a surprise.
Can you add an NSFW version?
You just need to use a NSFW base Checkpoint or loras, it works for me.
im getting nowhere near the quality of the results you got lol. using the workflow, didnt change any of the settings it comes with
If you want higher face consistency try setting the grounding_px to 1024, if you want stronger changes lower it to 384. All images in the examples were created at cfg1 with the turbo model and between 8 and 12 steps. :)
If it still doesn't work for you let me know and I can take a look at an example of yours if you want.
Can you share the workflow because there is no change at all with or without your "lora" on.
Hey, the workflow is in my GitHub. It's linked in the Lora description. You also need to download my custom nodes for the Lora to work :)
The workflow ships with krea2 raw as default, but the example images were actually created with turbo cfg1 and between 8-12 steps.
@conrad_locke hey, thanks your advice works!
@conrad_locke where exactly is that workflow?
@wieszjak here on my github. don't forget you also need to get my custom nodes. :)
https://github.com/lbouaraba/comfyui-krea2edit/tree/main/workflows
Another game changer. Works amazing. Gooners must be satisfied
Why can't I load your workflow on Git?
There are workflows on the Custom node github here: https://github.com/lbouaraba/comfyui-krea2edit
Man and woman in a pic. Can I use this to change only the man (and nothing else)? How? I see two prompt boxes.
I'm trying to use the two images wf.
If anyone can help me out, I'll tip. Thanks guys!
Did you try using the 1-image only wf and using a mask on the man?
The two prompt boxes are positive and negative conditioning. So when you are using the turbo model with cfg1 and 8 steps, you can leave the negative prompt box empty. Also my new 1.1 model should keep changes much more local, so changing one person should affect other elements of the image less. :)
@conrad_locke I saw your update, looks definitely better, thanks for your work!
Its great for recycling old material and create more variation of stuff you love. I used the single workflow on two people - the gag gallery picture. The orginal picture had them standing in row in the same perspective. I think it helps if you stay within the perpective corridor of the original image if 2 poeple are involved. Using her/him in the prompt will clone him/her once - if you use him/her/man/woman further in the pormpt, it will add more clones to the image and can give great clone effects - see braided dark tanned woman in gallery.Thanks for sharing i will use this to create lora datasets.
It helps describing them spatially and also mentioning a shared action like "looking at each other and laughing". That way they are generated in a more natural way. Also, sometimes just switching which person is input image1 and which person is input image2 helps. And thanks for your gallery submissions! :)
That works great! but if you using a character sheet it use the same character multiplied lol
That's a genuine gap in my dataset. I've got no examples of this and its a very useful capability. If you have example data of this, where you have the character sheet and the output image plus prompt, I would greatly appreciate it if you could upload them to your huggingface and notify me. That way I can include them in future training runs. :)
@conrad_locke would these help?
https://civitai.red/models/2761803/krea-2-subject-pose-sheet-by-astroburner
@conrad_locke I solved with the "the woman in image 1" like flux klein and anothers
@futer24769 Could you elaborate a bit? You put a character sheet in slot 1 and just referenced it as "the woman in image 1", and that worked?
v1.1 released!
General poll, what is most important for everyone next?
better face likeness
switch clothing
better multi image support
Inpainting model
Let me know :)
Face/body likeness is king for me...
face obviously
character consistency, so face/body likeness
likeness
Please do head swap
Regardless of the perspective, the consistency of the face/person is maintained.
Thanks for the incredible lora! Works even better when you add a background remover to the image (s) before inference. Also, if it's possible, please consider training in some character reference sheets, other than likeness, that would go a long way.
@ChevalierAI character consistency. Given How good Krea 2 Loras are Using this to create a high quality data set to train Loras. Would be quite useful.
For me? Face and body likeness first. Inpainting second.
face and prompt adherence, less blurry on generated image
This is exactly what I was looking for with Krea 2 til Krea 2 Edit... whenever it finally gets released. God knows when lmao, But why, oh why, is it 1.7 GB? 😭
Can someone please make it smaller?
I cant guarantee how well a lower rank version would work. but I can try to make one for you? :)
@conrad_locke That would be awesome, i will test it.
I have now uploaded a rank128 and rank64 version of the lora. so far to me it seems like they perform basically on par with the full version.
@conrad_locke tyvm, lemme test it, i don't see the files here tho,still uploading??,
i'm downlaoding from you hugginface page.
--edit--
it's on the dropdown menu.
@conrad_locke Rank 64 and 128 perform almost the same as the full model, but this cuts my generation time by around 60-70%. It's faster now, Thanks again.
fronyax, instead of giving a “thumbs down,” shouldn’t you be sharing the creative process instead? Unless you’re just trying to pull the wool over everyone’s eyes.
fronyax, instead of giving a “thumbs down,” shouldn’t you be sharing the creative process instead? Unless you’re just trying to pull the wool over everyone’s eyes.
@lighthorsexajz830 What are you smoking? 🤡, Imagine getting triggered by a tiny emoji and then stalking me. Be grateful the author shared this wonderful lora for free and was kind enough to share the workflow to the community in the first place. You're just too lazy to search for it yourself and would rather complain instead. Those thumbs down are well deserved.
🤡
@fronyax Share what? It's all just empty talk—the proof is in this post: there's nothing here. Where are the photos of this model that supposedly satisfied users, or where are their photos? There's nothing here—and it's certainly not the photo you posted; it doesn’t show anything and it’s really ugly—sorry, but you must have had a good laugh yourself posting that monstrosity :) Look at my post: I’m sharing without any problem. Just look at the stats. I didn’t create all my animations with your help—there’s a ton of work behind them—and you’re calling me lazy? :)
oooooh sexy a id likeness, one of reasons why we still fallback on klein9b. thnx!
just tried it with the workflows in lbouaraba/comfyui-krea2edit, with one of the example prompts with a 1:1 image. But cant get the likeness to work with the recommended setup.
Subject: Random padding behavior in v1.1 with different seeds
Hi!
First of all, thank you for the excellent work on Krea2 Identity Edit v1.1 — the face similarity and editing locality improvements are really impressive!
I'm experiencing an issue with random padding behavior that I'd like to report:
Problem:
When generating with v1.1 at 1024×1024 resolution with grounding_px=384-512, the model adds white padding around the generated image, but the padding position changes randomly with each seed, even though all other settings remain identical.
Observed behavior:
Same workflow, same settings, only seed changes
Padding appears sometimes on the left, sometimes on the right, sometimes centered
The generated content itself is good, but the random padding makes it unpredictable
My setup:
Model: krea2_identity_edit_v1_1.safetensors
Resolution: 1024×1024
grounding_px: 384-512
Input: portrait photo resized to square with crop
Questions:
Is this expected behavior for v1.1?
Is there a way to force consistent padding position (or no padding at all)?
Should I use v1.0 instead for stable rectangular output without padding?
Thanks for your help!
Additional finding (IMPORTANT):
I discovered that the random padding behavior only occurs when I use a cutout with white background as input.
When I use:
✗ Cutout (head only on white background via ImageAndMaskPreview) → Random padding appears
✓ Original photo with natural background → No padding, stable results
Why this matters:
Many workflows require working with cutouts (product photos, character isolation, etc.), so understanding how v1.1 handles white backgrounds would be valuable.
Questions:
Is v1.1 trained to interpret white background as "padding area"?
Should I avoid white backgrounds entirely, or is there a workaround?
Would using gray background (128,128,128) instead of white (255,255,255) help?
@soyv4 hey. Thanks for the finding! Can you link me the workflow you are using or ideally submit an image to the gallery that has the workflow embedded? That way I can take a look at it properly. Currently the model is trained completely without masks.
@conrad_locke Subject: Random padding in v1.1 — found the root cause (white background issue)
Hi!
First of all, thank you for the excellent work on Krea2 Identity Edit v1.1 — the face similarity and editing locality improvements are really impressive!
I'm reporting an issue with random padding behavior and I've found the root cause.
Problem:
When generating with v1.1 at 1024×1024 resolution, the model adds white padding around the generated image, and the padding position changes randomly with each seed, even though all other settings remain identical.
Root cause discovered:
The random padding only occurs when the input is a cutout with a white background (e.g., head isolated on white via ImageAndMaskPreview node).
Critical finding:
✗ Cutout on white background (255,255,255) → Random padding appears (left/right/center, depends on seed)
✓ Cutout on gray background (128,128,128) → No padding, completely stable results
✓ Original photo with natural background → No padding, stable results
This suggests that v1.1 interprets pure white pixels (normalized to +1.0) as "padding area to be filled randomly", while gray pixels (normalized to 0.0) are treated as neutral and filled with prompt context consistently.
My setup:
Model: krea2_identity_edit_v1_1.safetensors
Resolution: 1024×1024 (or 1024×1536)
grounding_px: 384-1024 (tried all values)
Input: head cutout from SAM3 + AGSoft Resize
Questions:
Is this expected behavior — does v1.1 treat white (255,255,255) as a special "padding token"?
Should users avoid pure white backgrounds when working with cutouts?
Could this be addressed in a future version (e.g., treating white as neutral like gray)?
Is there an official workaround for cutout workflows?
Suggested fix for users (workaround):
Replace white background (255,255,255) with gray (128,128,128) before feeding to Krea2 Edit — this completely eliminates the random padding issue.
I've attached examples showing the same settings with different seeds producing different padding positions on white background, and stable results on gray background.
Thanks for your work, and I hope this helps improve v1.2!
I haven't got much sleep tonight because of this lol
I'm happy you are enjoying the model! I love to see people's outputs, this immensely helps me improve the model further. So I would greatly appreciate it if you were willing to take the time to upload some of your generations and comment what you liked/disliked about them. :)
@conrad_locke I hope you continue to improve upon it. It has massive potential. Though today I had the problem making a portrait of a person. I used commands like "Create Portrait", "Transform the man into a portrait" among using a few chat GPT shenanigans but no dice. The person was taking a selfie while setting down. Probably too much to ask for lol
Are there Portraits in the training data?
Thanks! Is there a workflow involved?
It's in the github https://github.com/lbouaraba/comfyui-krea2edit
looks good, how does the 2 image work? does the node know image1 and image2 or is there some other reference scheme? Like give lady in image1 the hairstyle of lady in image2.
Currently there is no referencing the images through text. For example in order to place a character in a scene, i recommend to place the scene as image1 and the character as image2 because that's the order the model is currently trained with.
Wow man this is really awesome. Krea must hire you at this point :D
What was your settings ? How many steps and images?
Thank you! what are the r64 and r128 files for?
They are smaller so better for people with less VRAM, though at a very slight quality drop.
The width and height should be divisible by 8; otherwise, there will be ghosting around the details.
I was getting this, thanks for the explanation
Thank you for figuring this out and letting people know! :)
I had the ghosting issue, too, even with height and width divisible by 8. I could only get rid of it when I made it divisible by 64, like in 1280x1088.
@eladamri The specific reason is the misalignment between the aspect ratio of the image tensor fed to the model and the input latent. The latent input should automatically align with the model's requirements. Try applying aspect ratio constraints as the first step in the image output process. Try using comfy-rmbg's image resize function, which automatically ensures that the aspect ratio is divisible by integers. Then, feed the adjusted aspect ratio to the empty latent generator and other latent generators. This should solve the problem.
@Lemonton Thanks, that's exactly what I did. I just wanted to mention that at least for me, it wasn't enough to make width and height divisible by 8 - the issues only disappeared when I made them divisible by 64. Maybe others make a similar experience. Or maybe that's just me.
@eladamri Tx, had the same problem and only this divisible by 64 corrected this.
Here is s small list of common resolutions that should work :
576 × 1024
704 × 1280
768 × 1344
832 × 1472
896 × 1600
960 × 1728
1024 × 1792
1088 × 1920
1152 × 2048
@JimTheKing maybe it's each side divisble by 8, that makes it 64. not sure how the resize node works
Hii i'm interested in trying this out, is there a workflow somewhere? I tried the images you posted but they don't contain it.
nevermind i found it https://github.com/lbouaraba/comfyui-krea2edit/tree/main/workflows
That's really too bad. It doesn't make any sense.
Welp. There went my entire afternoon playing around with this thing. Super fun and very useful.
Thanks! New updates are coming soon to increase quality further :)
This is an excellent workflow and LoRA combination. However, I would prefer if the author merged some of the KREA2 NSFW LoRAs (apologies, I am an NSFW creator myself); this could reduce VRAM usage associated with loading multiple LoRAs. Also, due to the training dataset, some non-standard output dimensions—specifically those violating the maximum edge length used during training—result in a loss of character likeness. Overall, it is a fantastic workflow and LoRA; I hope the author continues to update it. KREA2 has great potential.
How do I use this with SwarmUI?
Download the lora, open the SwarmUI folder, go to models, put it in the lora file
Is it just me or is this censored?
i think it depend on your krea 2model/checkpoint...
Thanks for sharing, but I'm not interested—where's the workflow?? Your example images don't include any data, so this won't be much use except to look elsewhere. It's really a shame.
There's a template included with the Krea2Edit extension needed for this.
Thanks for your feedback, but I already have Krea2 Identity Edit—unless that’s not the same thing? But my criticism of the person who shared their Lora file is that they didn’t provide even a basic workflow to give us time to download any missing extensions or models, and this not seem to be the case, because once I figure out that workflow myself, I’d share it to round out the whole thing—it’s just common sense. I’ve adopted this approach for all the models I’ve created myself; otherwise, how do you expect other users to share their work? The proof is right here—I’m not blind; you can see from this post that there’s no feedback at all in terms of sharing :)
hope you can make this lora run more faster ..anyway its awesome btw..thx
You're a hero, nice work!
Has anyone tried to use this lora for generating "instagram carousel" photos? same room different angles/expressions etc
Has anyone tried to use this lora for generating "instagram carousel" photos? same room different angles/expressions etc
This is incredible lora that turn Krea 2 into model that to have edit function. It really impressive and a true edit function. The inference time is significantly longer , Thanks you very much for the idea and effort. i think this lora worth more attention.
Thank you! I would love to see your generations :)
@conrad_locke just posted some side by side image.