Krea 2 Identity Edit

The trace is too big to post as a comment here, but here's claude's analysis:

## What the exception is

The top-level error is:

torch.AcceleratorError: CUDA error: invalid argument (cudaErrorInvalidValue)

It was raised on node 53 (KSampler), but the crash point is a tensor.to(device) copy - a CUDA memory operation, not a math kernel.

## Where it actually happened

Reading the stack from bottom to top, this is NOT a normal KSampler run. A custom node injected itself into the diffusion model's forward pass:

1. KSampler -> sample_euler -> diffusion model forward (krea2/model.py)

2. A DIFFUSION_MODEL wrapper from custom_nodes\comfyui-krea2edit\__init__.py:304 runs inside the denoise step

3. That wrapper calls fitencode_image(...) -> vae.encode(...) (line 114)

4. vae.encode calls model_management.load_models_gpu(...) to make room for the VAE

5. To free VRAM it does free_memory -> partially_unload -> m.to(offload_device)

6. The weight-transfer t.to(device, ...) throws CUDA error: invalid argument

So the immediate trigger is: the krea2edit node runs a VAE encode in the middle of the sampler's forward pass, which forces ComfyUI to swap models on the GPU, and the offload copy fails.

## Likely root cause

- The stacktrace is probably misleading. ComfyUI even prints: "CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect." cudaErrorInvalidValue surfacing on a plain .to() almost always means an earlier async CUDA kernel already faulted, and the error only became visible here. The real failure is likely upstream (e.g. inside the compiled krea2edit encode path or the DiT forward), not the memory copy itself.

- Aggressive memory setup + torch.compile. Running with --vram-headroom 1.0 (only 1 GB headroom) on the cudaMallocAsync allocator, with torch.compile/Dynamo active (there is a WON'T CONVERT warning right before the crash). Doing a VAE encode + model offload/reload during a compiled forward pass, with almost no free VRAM (log shows ~3.7 MB Torch VRAM free), is a fragile combination that can corrupt the allocator/context.

## How to narrow it down / fix

1. Get the true error location. Relaunch with CUDA_LAUNCH_BLOCKING=1 (and ideally TORCH_USE_CUDA_DSA). This makes CUDA report the failing kernel synchronously so the stacktrace points at the real culprit instead of the .to() copy.

2. Disable torch.compile for this run to rule out a Dynamo/compiled-kernel interaction (krea2/model.py uses torch._dynamo).

3. Loosen memory pressure. Increase --vram-headroom (e.g. 2.0+) or try the default (non-async) allocator by setting PYTORCH_CUDA_ALLOC_CONF without cudaMallocAsync, so the mid-forward VAE encode does not force a risky partial unload.

4. Suspect the custom node. The whole crash flows through comfyui-krea2edit, which does an unusual vae.encode inside the DiT forward. Try the same workflow with that node bypassed/removed - if it succeeds, the node's mid-forward encode + offload is the problem.

The most useful next step is running with CUDA_LAUNCH_BLOCKING=1, since the current trace is very likely pointing at a symptom rather than the actual failing operation.

Removing torch compile seems to help, but do the k2edit nodes really need to to the vae encode inside the forward pass? Can't this be done outside once and passed in?

https://github.com/lbouaraba/comfyui-krea2edit/tree/main/workflows

@2legsRises357 Thanks so much!!

What would be the difference of making it rank 128/rank 64? I know I can google this but an actual human perspective is always better.

@ProvenFlawless Less memory usage for me, and faster gen time.

You can try INT8 version. it's much more time saving. i'm using Qwen 3VL_4b_int8.safetensor

exact same error here, right click the textencode node and reload or recreate, it works then but without the fidelity booster and makes the editing useless

@ksgaming87726 well, OP doesn't seem to care so whatever :)

I have. still working on figuring it out. i used a reference of myself and it its a portrait upper torso shot on landscape, face is still not detailed. experimenting with creating an inpaint to fix it

klein 2 already obsolete.
For most edit this model already nice, for hard edits - big love qwen. For generation - krea 2 with vantage condition nodes (more than twice fast).

@conrad_locke np, it was a commit to comfy/ldm/krea2/model.py had to revert to stable to get Krea2 edit working again...

https://github.com/Comfy-Org/ComfyUI/commit/c9602625e445e9ee37d3ac6faf5ea9ec1e0de87e

I keep my Comfy updated, this was the only change I could see that would have messed with how Krea2 worked.

@conrad_locke Thank you for the quick update!!

you just need to write proper prompt to.. let say, "Replace the face, hair and head of main character dressed in blue with the girl in image 2. The original outfit should remain the same".

Came here to ask for this. I can't seam to make it work.

it work with my character lora. didn't see any problem.

@ianklpoon347 Could you give more insight on how? You just added another lora loader after the identity edit lora? Any other settings? How did you prompt it?

@lost_moon i used ND Super Lora Loader node (I find this to be the best loader) and have any loras + identity edit lora loaded in there and it seems to work.

@kresl it works when I add the TextFusion Refusal-Reduction lora. 👍

@lost_moon I tried with the filter bypass lora and also the TextFusion Refusal Reduction lora, but still my other (n s f w) loras won't have any effect. Can you help in any way?

@dachhaus41200 you might need to prompt it differently, always keep in mind you need to reference "image 1" and might need to change weights too. it's not easy.

@lost_moon Thank you, I will try again.

Yes, make a mask.

why dont you just do it?

Lmao, skill issue.

"Right, skill issue. My 1400+ videos, 4.2k followers, 740 likes, and 8.3k downloads are definitely the portfolio of someone with no clue what they're doing. Thanks thochapr for diagnosing my competence based on one feature hiccup — I'll go delete all those thousands of downloads that clearly happened by accident. Super helpful feedback. 👏"

@thochapr "'Skill issue' — coming from someone with 3 images that hurt to look at and 3 followers who probably followed by accident. Meanwhile I'm sitting here with 1400+ videos, 8.3k downloads, and an actual community. But sure, I'M the one lacking skills. Thanks for the wisdom, thochapr — once you figure out how to create something watchable, we can talk again. 💀"

"21 ghost likes on a pointless comment. Even bots are embarrassed to engage. 😂"

Looks like you may need to update your nodes gergarding this LoRA.

@Iridu no, I only see this message with the latest version of ComfyUI. When I revert to an older version of ComfyUI, the error no longer occurs.

Yes, such a stupid guy sounds not even know how to use this lora AND don't know the purpose of using this lora.

@lizardon1024 Thanks for the passionate feedback! I have to admit, I love it when someone calls me an “idiot with an ulterior motive” without even sharing a single result of their own work. It’s like a chef criticizing Michelin without ever publishing a recipe. But anyway, I’m sure your results with Identity Edit are amazing—I’d love to see them someday. Wait, you’ve never posted them? Oh, what a shame! 🎭

@ianklpoon347  @ianklpoon347 “Stupid guy” — you know what's really stupid? Having the nerve to criticize someone WITHOUT ANY IMAGES to show. That's literally the opposite of credibility. But go ahead, show me your masterpieces with that LoRA so we can laugh about it together. Oh, wait... you don't have anything? Oh, okay. 😂

@lighthorsexajz830 hijacking this because your review got locked and I had a feeling you'd be replying here: you're crashing out over people RIGHTLY pointing out you fucked up your config and acted like the model was at fault for it.
My dude, even discarding the fact that you are acting like you're a big deal over a few people looking at your stuff - which, honestly, is not even good (jfc,"a chef criticizing a Michelin", talk about an unwarranted sense of superiority) - the person you replied to might not have shown any images, but other people have. What are you implying here, that they're all faking their results?
No one is even saying that Krea with this lora is as good as other image edit models, just that YOUR results are not demonstrative of the model working as intended. Again, as I said before: you're acting stupid. Have the tiniest bit of self-awareness, leave your narcissism and unwarranted self-worth at the door, and stop acting stupid.

EDIT: For the record, I don't think you have any kind of agenda. You're just so damn full of yourself that you can't accept fault for misusing the model.

"...that even outperforms 𝙵̸𝚕̸𝚞̸𝚡̸ ̸2̸ ̸𝙺̸𝚕̸𝚎̸𝚒̸𝚗̸ Nano Banana in many aspects."

Fixed ;)

@lighthorsexajz830 
People usually use editing models for private images, so most people (like me) don't upload their results. But since you keep pestering everyone with your embarrassing inability to use this LoRA properly, here are some examples:
Original image: https://civitai.red/images/106854812
Hair color change: https://civitai.red/images/137457798
Dress change: https://civitai.red/images/137457799
Hair style change: https://civitai.red/images/137457801
No seed adjustments, simple prompts, upscaled via SeedVR2. The workflow is embedded into the images. Yes, it's really that simple. Now go and delete your pitiful "tests"...

Good evening or good morning, depending on where you are! :)

It's been a few hours since I downloaded this LoRA along with the required nodes, and I'm still experimenting with it.

I can confirm that the anime-to-realistic conversion you're asking for is definitely possible, especially since I'm a big fan of this kind of transition.

I’m using Ollama Vision for automatic prompt conversion, trying to find the 'golden ratio' (the perfect sweet spot). (I know it's not really necessary for this).

You can write something like:

Transform this image into a highly realistic, raw photograph. Make sure the subject has true-to-life proportions, natural structure, and absolutely NO anime, cartoon, or stylized features. Avoid big eyes and smooth plastic skin. Include realistic skin texture with pores and slight natural imperfections. Use natural, cinematic lighting. It should look like it was taken with a DSLR camera. Maintain the original color palette of the uploaded image, but ground it in reality

@pantelis1985 I tried your suggested universal prompt on different Anime images and it's not entirely impossible, but from my own tests, Turbo usually fails to convert it to realism for everything except the face. Raw works a little better but the conversion often changes more than necessary and then has quite bad adherence to the original image. It will probably work a little better with a custom prompt for each image (like when using llm vision models) and with an additional realism LoRA, but Flux2 Klein with a simple "Change image to a realistic photograph." prompt is so much more convenient and still works better in most cases. So yeah, a feature like that in Krea 2 Identity Edit would be highly appreciated. :)

@lizardon1024 @pantelis1985 
Krea2 Transformation Lora
+
Krea2 Realism Slider
=
💸

@strash The transformation LoRA seems to be for entirely different purposes and the realism LoRA helps a little bit, but at least my results are still far worse than Flux 2 Klein 9B with a simple "Change image to a realistic photograph." prompt... But still thank you for suggesting those LoRAs!

install the lastest version of the node, I have the same error before MAJ, now it's OK

Stop trying to be conspicuous, bro. No one really cares if it suits you or not.

i'm also using krea 2 turbo, what alternatives to this edit workflow do you prefer? Thanks. My low VRAM is showing :(

@h80857248984 hey jackal 😏 “So you download my content, you steal it, and you have the nerve to criticize me? That’s great. Let me know when you’ve created something on your own. In the meantime, thanks for the free publicity! 😏”

following because im wondering the same

Unfortunately, this is probably normal. The model is larger than your vram, which comfy handles well, but when you add higher precision Loras, on-the-fly quants conversion as comfy move the model between RAM and VRAM slows down significantly. I guess the only solution is more VRAM that will fit the entire model

I think it's not about the size, but about the fact of changing quantization when patching from bf16/fp16 to fp8. this problem occurs when using almost any Loras. I'm curious why gurupatrik is writing about this now, as if he didn't have this problem before.

EDIT: lol it works.

you can have two input.