Krea 2 Cannot Edit Images. Here Is How People Are Teaching It To Anyway
A training flag, two ComfyUI nodes, and a folder of before-and-after pairs turn a text-only model into one that takes a picture as its instruction.
Krea 2 has no image editing. You cannot hand it a photograph and ask for a change. It reads a sentence and paints from nothing, which is what it was built for and part of why people like how it looks: the model was tuned for aesthetics rather than obedience, and it shows.
Then two edit add-ons appeared on Hugging Face inside twenty-four hours, both for Krea 2, both taking a reference picture as input. One turns a line drawing into a finished image. The other builds a four-view character turnaround from a single reference. Neither of them should have been possible on a model with no reference path at all.
The path is new, it is two weeks old at most, and it is worth understanding before you spend an afternoon on it, because the reason it works is also the reason three different things can fail without telling you.
Why a picture beats a sentence
Every prompt you write is a translation loss. You have an exact colour in mind and you type "warm ochre." You have a specific garment and you type "cropped jacket." The model picks from everything those words could mean, and you spend the next forty minutes narrowing it with adjectives.
A reference picture removes the translation. You are not describing the jacket, you are handing over the jacket. Anyone who has worked with an art director knows the difference between a note and a marked-up frame, and this is that difference.
The catch is that a model has to be built to accept one, and Krea 2 was not.
What was added, and where
Two pieces, both by ostris, the author of ai-toolkit, the training tool that most independent add-ons for these models come out of. It has 12,000 stars.
The first is a training flag. The krea2 architecture in ai-toolkit now accepts model_kwargs.edit: true, and turning it on builds a reference-image path into the training run. With the flag off, which is the default, none of it exists and you are training an ordinary style add-on.
The second is ComfyUI-Krea2-Ostris-Edit, a two-node pack with no extra dependencies at all, 329 stars. It is what lets you actually run what you trained. Its own one-line description calls the training path experimental, which is the right word and worth carrying with you through the rest of this.
Two adapters already exist and both landed on September 14. Cierpliwy/krea2_lineart_edit, published at 20:37 UTC with 8 likes, does drawing to finished image and its four-line card says it works best against a realistic lineart preprocessor. WarmBloodAban/Krea-Edit-lora, at 08:11 UTC, builds four-view character sheets from text or from a reference, and ships its ComfyUI graph inside the repository.
The mechanism, which is genuinely clever
Your reference picture goes into the model twice, at two different sizes, for two different jobs. This is the part worth understanding, because every trap below comes out of it.
Job one: understanding. The picture is shrunk to fit inside roughly 384 by 384 pixels of total area and fed through the model's language side, sitting next to your prompt. That side is a vision-and-language model, and it does not need detail. It needs to know that there is a person, facing left, in a red coat, on a beach. A thumbnail tells it that. Sending a 12-megapixel file would tell it exactly the same thing at forty times the cost.
Job two: fidelity. Separately, the same picture is compressed into the model's internal image format at up to a full megapixel by default, and those compressed frames are appended to the picture the model is painting. They sit there from the very first pass, clean, never scrambled with noise the way the target is. The model only ever predicts the target. The reference is context, permanently sharp, sitting in the same visual sequence.
So the language side gets the gist and the image side gets the pixels, and neither is asked to do the other's job. That split is the whole design, and it explains a behaviour people find surprising on first use: your reference keeps its own shape. It is never squashed to match your output. A wide reference and a tall render work together, because the reference was never treated as a canvas in the first place.
Up to three references at once, which is what makes the character-sheet use case work.
Put this into practice
This is a training job, not a prompt trick. Here is the honest shape of it.
Build the pairs. You need a folder of finished results and a matching folder of the thing you started from, with identical filenames. In ai-toolkit's language, folder_path is the target and control_path is your reference. Caption each target with a plain text file of the same name. The size and aspect ratio of the reference does not have to match the target, which is unusual and useful.
Be specific about what you are teaching. "Edit images" is not a task. "Turn my studio line drawing into a finished render with the linework intact" is a task, and it is the kind of thing a folder of thirty pairs can actually teach. The two adapters that exist both picked one narrow behaviour.
Write the config. There is no Krea 2 edit example in ai-toolkit's examples folder, checked this morning on the main branch. Nobody wrote one yet. Start from train_lora_qwen_image_edit_2509_32gb.yaml, which has the same paired-folder shape, and change the architecture to krea2 with model_kwargs.edit: true. Leave the kv_cache option alone; see below. That template's own defaults give you a sense of the scale involved: 3,000 training steps, one image at a time, a small add-on file rather than a full retrain. Expect hours, not minutes, on one rented card.
Wire it up in ComfyUI. Clone the node pack into custom_nodes and restart. The chain is: load the Krea 2 model, load your add-on, run it through Krea 2 Ostris Edit Model Patch, then into the sampler. On the prompt side, Text Encode Krea 2 Ostris Edit takes your text, up to three reference images, and a connection to the image compressor. That node is safe to leave in a graph permanently; with no reference attached, it behaves exactly like the stock model.
Cheaper first step. If training from scratch is more than you want to commit to today, install the node pack and download krea2_lineart_edit, which is a one-file download. Feed it a drawing. You will know in ten minutes whether reference-image editing on this model is worth your afternoon. There is also a hosted Krea 2 add-on trainer at multimodalart/krea2-lora-trainer if you would rather not stand up a training box, though it is a general trainer and I have not confirmed it exposes the edit flag.
Three ways this fails without telling you
I want to be direct about these, because all three produce a finished image. None of them errors.
The patch node is not optional. Stock ComfyUI's Krea 2 loader ignores reference pictures completely. Not partially, completely. Without Krea 2 Ostris Edit Model Patch in the chain, you get a perfectly good text-to-image render that has never seen your reference, and nothing in the console says so. If your first result looks like the model ignored your input, this is why, and it is the single most likely cause.
The language side has to have its eyes. The text encoder file you load must include the vision half of the model. The node pack's README states this outright, and the reason it is worth stating is that dropping the vision half is a reasonable thing for a plain text-to-image build to do: it is dead weight taking up memory when nothing is ever going to show it a picture. Load one of those and your reference pictures cannot be read at all.
The speed option is a correctness trap. The patch node has a kv_cache toggle, off by default. Turned on, it does something sensible: it works out the reference's contribution once at the start and reuses it on every pass, instead of dragging it through every step. On a long render that is real time saved.
It only works if the add-on was trained with the matching option on. The base Krea 2 model was trained with everything looking at everything, so the reference's contribution genuinely does change from step to step unless training specifically taught it not to. Flip the toggle on a normally trained add-on and the model reuses a value that was never stable, and you get an image. A wrong one, silently. Leave it off unless you trained it on, and if you are downloading someone else's add-on, assume off, because neither of the two published adapters says a word about it.
One more thing about the licence
Krea 2 Raw is gated behind a click-through and governed by the Krea 2 Community License, with a separate acceptable use policy. Read both before you build anything commercial on this.
And check the tags on anything you download. WarmBloodAban/Krea-Edit-lora declares Apache 2.0 on its card. The model it modifies is not Apache 2.0, and an add-on cannot relicense the thing it needs in order to run. The upstream terms are the ones that govern what you make. That is not a knock on the person who published it; metadata fields get filled in fast and nobody is checking. It is a reason to read the base model's licence rather than the adapter's.
The gap worth filling
Here is the state of this, plainly. The capability is real, the mechanism is well designed, the tooling works, and the two adapters that exist are hobby work with no training details, no evaluation, and in one case a licence field that is wrong. Nobody has published a Krea 2 edit adapter with a documented dataset, a step count, a sample grid, and a note on whether the speed option was enabled.
That is not a complaint. That is an opening. The recipe above is the whole recipe, the tooling is free, and the first person to publish a properly documented edit adapter for the best-looking open image model available will have made something a lot of people want.
If you have thirty before-and-after pairs sitting in a project folder from a job you already finished, you have the hard part already.
Medium metadata
Title: Krea 2 Cannot Edit Images. Here Is How People Are Teaching It To Anyway
Subtitle: A training flag, two ComfyUI nodes, and a folder of before-and-after pairs turn a text-only model into one that takes a picture as its instruction.
Tags: AI Art, ComfyUI, Generative AI, Creative Tools, Machine Learning
Suggested read time: 8 minutes