What Krea 2 img2img actually does: it re-renders, it does not edit
Krea 2 img2img does not edit your image. It reads the picture with a vision-language model, turns what it sees into a description, and generates a brand-new image from that description plus your prompt. Palette, lighting, composition and mood carry across. The exact face, the exact product, the exact logo and the exact text do not.
That one fact explains almost every complaint about Krea 2 image to image: the style lands, the subject does not survive. If your job is "make this look different", Krea 2 is a strong choice. If your job is "change this one thing and leave the rest alone", you need an editing model instead — that is what Img2Img AI is built for.
The evidence is in Krea's own documentation. Krea's technical report states that Qwen 3 VL is the model's text encoder, picked over T5-XXL for its richer text-and-image input space. ComfyUI's official Krea-2 tutorial describes the model plainly as text-to-image, shipped as two checkpoints: RAW at 52 steps for training, Turbo at 8 steps for inference. Neither page describes an edit mode, because there isn't one.
What Krea 2 is built to do
Krea 2 is Krea AI's own foundation image model. It launched inside Krea's app in May 2026 and went open-weight in June. It was trained for aesthetic range and style control rather than literal prompt obedience, and that priority shows up in every control it exposes.
| Control | What it does | What it means for img2img |
|---|---|---|
| Style references | One or two images steer the look, each with its own strength | Krea's real image input — it transfers style, never content |
| Moodboards | A wider set of visuals sets creative direction | Direction-setting, not per-image control |
| Creativity | How far the model may interpret what you left unsaid | Higher creativity means more drift from your source |
| Style LoRAs | Nine official trigger-word LoRAs, documented at strength 1.0 | Medium and texture, applied globally |
| Open weights | RAW at 52 steps for training, Turbo at 8 steps for inference | Lets you build img2img yourself in ComfyUI |
| API and app access | Hosted inference, no local setup | Same behaviour, different billing |
Read that list again and notice what is missing: nothing in it preserves a subject. Style references carry a look. Moodboards carry a direction. Creativity actively loosens the model's grip on your input. There is no control anywhere in Krea 2 whose job is "leave this part of the image alone", and that absence is the through-line for everything below.
Can Krea 2 do image to image? Yes, through three different doors
"Can Krea 2 do image to image" is the most common version of this question, and the honest answer is that three separate routes all get called img2img, and they behave nothing alike.
| Route | What it is | What it preserves | What it needs |
|---|---|---|---|
| Style reference | Krea's first-party feature: one or two reference images steer style, with continuous strength per reference | Palette, medium, lighting, mood | Krea web app, or the Int8 style-reference workflow in ComfyUI |
| Classic img2img | VAE-encode your image, re-noise it, sample back down | Rough composition and colour blocking | Community ComfyUI workflows |
| Identity Edit LoRA | A community fine-tune plus a custom node pack that feeds your image in as both latent and semantic conditioning | Faces, objects, untouched regions | ComfyUI plus the Krea2Edit node pack — see Krea 2 image editing |
Only the third route behaves like editing, and it is community work, not a Krea feature. The first two are restyling with extra steps.
Image to image vs image editing: the one difference that breaks most attempts
An editing model is judged on what it leaves alone. An img2img pass is judged on what it changes. That is the whole difference, and it decides which tool is right before you touch a single slider.
Run a portrait through Krea 2 img2img and you get a portrait in the new style, of a person who looks like your subject's cousin. Run a product shot through it and you get the same bottle with a slightly different label, a different cap thread, and a brand colour that is close but not your hex. Run anything with words in frame through it and the words come back as convincing nonsense.
None of that is a bug. The model was asked to generate an image, and it did. It was never asked to preserve pixels, because Krea 2 has no mechanism that preserves pixels.
Strength values that hold up (and where the face falls apart)
Krea has not published a denoise curve for img2img, because Krea does not ship a first-party img2img mode outside style reference. Every number below comes from ComfyUI's official docs, the LoRA author's own model card, or community setup guides, and every one of them is hardware- and image-dependent. Treat them as starting points, not settings to copy blindly.
Sampling, from the ComfyUI Krea-2 tutorial
- Turbo defaults to 8 steps with prompt enhancement on and no LoRA.
- Krea 2 Turbo is distilled to run at CFG 0.0. It does not want the 7–15 CFG range you may be used to; community guides settle around CFG 1.0 for general use and roughly 12 steps at CFG 2.5 for img2img passes.
- Output resolution runs from 1K to 2K. Set the ResolutionSelector to 2.0 megapixels for 2K.
- Krea's nine official style LoRAs are all documented at strength 1.0, each with its own trigger word.
Identity Edit LoRA, from the author's v1.1 notes
- Turbo at 8–12 steps, CFG 1.0 for most edits.
- RAW at 20 steps, CFG 3.0 when you are removing something.
grounding_pxbetween 384 and 768, default 768. Lower values push harder on the edit, higher values protect identity. That dial is the whole edit-versus-likeness tradeoff in one number.- Keep output at or below 2MP, and drop to 1–1.5MP when two people are in frame, then upscale.
The practical read: below roughly 0.4 denoise you get a slightly cleaner version of the same picture, above roughly 0.6 you get a different person wearing your subject's outfit, and the useful window in between is narrow enough that most people give up and reach for an editor.
What Krea 2 img2img is genuinely good at
- Medium flips. Photo to anime, anime to photoreal, sketch to painting. The subject drifting is acceptable here because the subject is not the point. If that is your whole job, a browser tool like cartoon to realistic AI or photo to illustration does it in one upload with no checkpoint to download.
- Style transfer from references. One or two reference images, continuous strength per reference. This is the feature Krea actually built, and it is good.
- Concept exploration. Feed it a rough draft, get ten directions back. Aesthetic range is what Krea 2 was trained for.
- Moodboard-led art direction. Establishing a look before anyone commits to a final asset.
Where Krea 2 img2img breaks: faces, products, text, logos
- Faces. Likeness drifts on every pass. If the job is swapping or keeping a specific face, use a face swap tool instead of fighting denoise values.
- Products and brand colour. Regeneration re-invents labels, materials and hex values. For a controlled colour change, AI recolor changes the colour and nothing else.
- Text and logos in frame. Re-rendered text comes back garbled. Krea 2 has no mechanism to leave a region untouched.
- Outpainting and inpainting. Krea 2 ships neither. Extending a frame needs a dedicated AI image extender; repairing part of a photo needs a real inpainting pass, the way old photo restoration does it.
If you are running Krea 2 locally and want the checkpoint details, step counts and VRAM numbers, that is all in the Krea 2 Turbo guide.
Four mistakes that make Krea 2 img2img look broken
Treating it as an editor. The most common one, and the one no setting fixes. If the sentence in your head starts with "keep everything but…", you are describing an edit, and Krea 2 does not do edits.
Bringing Stable Diffusion CFG habits. Krea 2 Turbo is distilled at CFG 0.0. Feeding it the 7–15 range that worked on older checkpoints produces burnt, oversaturated output, and people read that as the model being bad.
Pushing denoise up to force the change. When a restyle is not landing, the instinct is to raise denoise. That does make the change land — by discarding more of your source, which is exactly the subject you were trying to keep. The fix is a stronger prompt or a style reference, not a higher denoise.
Adding the style reference last. References set direction. Bolt one on after the composition is locked and you get a fight between two visual languages. Set the reference first, then work the prompt.
Krea 2 image to image vs Img2Img AI
| Krea 2 img2img | Img2Img AI | |
|---|---|---|
| Core job | Re-render in a new style | Change one thing, keep the rest |
| Keeps the subject's face | No | Yes |
| Keeps product colour and label | No | Yes |
| Outpainting and inpainting | Not supported | Built in |
| Setup | Krea account, or ComfyUI plus a 13–26 GB checkpoint | Browser, upload, done |
| Time to first result | Minutes to hours if you install locally | Seconds |
| Best for | Art direction and style exploration | Production edits on images you already have |
The verdict, with no hedging: if you are exploring a look, use Krea 2. If you already have the image and need a specific change to it, Krea 2 is the wrong tool and no setting will fix that. Use an editor.
Doing the same job in a browser, step by step
- Open Img2Img AI and upload your source image.
- Write the change as an instruction, not a description. "Change the jacket to olive green" beats "a person in an olive green jacket".
- Name what must not move — the face, the logo, the background, the text.
- Generate, then refine on the result rather than starting over.
The whole loop takes seconds, runs on any machine, and keeps the subject you started with.
Quick answers about Krea 2 img2img
Can Krea 2 do image to image?
Yes, but not as an edit. Krea 2 describes your image with a vision-language model and generates a new one from that description. Style carries over; the exact subject does not.
Does Krea 2 have a denoise or strength slider?
Not as a first-party control. Krea's own image-to-image feature is style reference, which has continuous strength per reference image. Classic denoise sliders only appear in community ComfyUI workflows.
Why does Krea 2 change my subject's face?
Because it is generating a new image rather than editing yours. Nothing in the pipeline is asked to preserve pixels. The community Identity Edit LoRA is the only route that addresses this, and it needs a custom node pack.
Can Krea 2 edit an existing photo?
Not natively. The closest option is the community Identity Edit LoRA — how it works and the settings that hold up are covered in can Krea 2 edit images.
What is the best Krea 2 img2img workflow in ComfyUI?
Start from the official Krea-2 Turbo template, then add either the Int8 style-reference workflow or the Krea2Edit node pack depending on whether you want restyling or editing. Model files, VRAM and step settings are in the Krea 2 Turbo guide.
Does Krea 2 support outpainting or inpainting?
No. Neither is part of the model. Use a tool with a real outpainting pass, such as AI image extender.
Krea 2 or Z-Image Turbo for image to image?
Different tradeoffs. Krea 2 wins on aesthetic range and style reference; Z-Image Turbo is faster, lighter and better at keeping text readable. Neither is an edit model.
Try an image-to-image edit that keeps your subject
If your interest in Krea 2 img2img is really about turning an image you already have into a cleaner, more controlled version, skip the checkpoint download. Upload the image to Img2Img AI, say what changes, say what stays, and keep the face you started with.
