Krea 2 Edit Review: Downloads, ComfyUI Workflows and Prompt Examples

Krea 2 Identity Edit is worth trying when you like Krea 2’s look but need to change an existing character, product or scene. It adds instruction-based editing to a local Krea 2 setup through a LoRA and custom nodes. Its appeal is keeping your existing model and LoRA workflow; its limitation is that identity and untouched pixels are not guaranteed to stay exact.

Here, “krea-2-edit” means conradlocke’s unofficial Krea 2 Identity Edit community project. It is distinct from Krea’s hosted Edit feature and should not be presented as a separately released official Krea editing checkpoint. Existing Krea 2 users can go straight to the downloads and starter settings.

Information checked: September 28, 2026.

What Krea 2 Edit actually adds

Version 1.2 extends the editing feature set with reference sheets, try-on, inpainting and outpainting, alongside likeness improvements. Node version 1.2.5 is a runtime update, not a new set of weights. Keeping those two version numbers separate avoids unnecessary downloads. See the maintainer’s changelog.

Model, LoRA and workflow downloads

If Krea 2 already runs on your machine, reuse compatible base-model, encoder and VAE files. Check the author’s file list and the Comfy Org repackaged files before downloading duplicates.

The full LoRA is about 1.83GB. File size is not the same as inference VRAM. Select the smaller filename in the LoRA loader if you use r128/r64. For the Raw recipe, obtain compatible weights through Krea’s Raw model repository or the diffusion_models folder in the Comfy Org repository.

Your first ComfyUI edit, step by step

1. Install the nodes and load the supplied JSON

cd ComfyUI/custom_nodes
git clone https://github.com/lbouaraba/comfyui-krea2edit

1. Use a ComfyUI build with native Krea 2 support, add the node pack, and restart. Do not clone over an existing folder of the same name.
2. Save and drag the author’s JSON onto the canvas. If nodes are red or missing, check their names and your installed versions first.
3. Select your actual files in the model, CLIP, VAE and LoRA loaders. Set the CLIPLoader type to krea2.
4. Leave the workflow in single-image mode and load your own source. You do not need the placeholder example.png.
5. Request one change, fix the seed and generate one image. Compare it side by side with the source before increasing complexity.

2. Keep both reference paths connected

This is more than adding a LoRA to ordinary img2img. The source must reach both the model’s reference-latent path and the image-grounded text encoder. Replacing Krea2EditGroundedEncode with a plain CLIPTextEncode removes the semantic image path.

LoadImage → VAEEncode → Krea2EditModelPatch.source_latent
LoadImage → Krea2EditGroundedEncode.image
UNETLoader → LoraLoaderModelOnly → Krea2EditModelPatch → KSampler.model
Krea2EditGroundedEncode → KSampler.positive
EmptySD3LatentImage → KSampler.latent_image + Krea2EditModelPatch.target_latent
KSampler → VAEDecode → SaveImage

For fit mode, also connect vae and source_image on the patch. Connect the same output latent to target_latent and KSampler. This lets source encoding happen before sampling instead of triggering model offload mid-run on VRAM-tight systems. Node wiring documentation.

3. Change the recipe to match the edit

The model card suggests 8–12 steps for Turbo and 20 / CFG 3 for Raw removals; a note inside the JSON also suggests 40 / CFG 3–4 for Raw. These are starting points, not a single mandatory preset. When CFG exceeds 1, use an image-grounded negative with an empty prompt and the same image.

4. Move on to scene-plus-person editing

Group 2 ships bypassed. Right-click its title and turn Bypass off. Put the scene on image/source_latent and the person on image_b/source_latent_b. Start with one person and an empty scene; two-reference support is not a promise of reliable two-person face separation.

Eight use cases and prompt examples

These are original starter prompts to adapt, not prompts tied to measured outputs in this review. Keep the first pass short: action, target, then the most important things to preserve. Add details only when a specific failure calls for them.

Product recolor

Change only the red fabric panels of the sneaker to cobalt blue. Keep the sole, stitching, logo, camera angle and background unchanged.

Start with Turbo; inspect logos and seams at full size. A convincing preview is not an exact product specification.

Relight a portrait

Relight this portrait with soft window light from the left. Keep the same face, expression, hairstyle, clothes and framing.

Compare facial proportions as well as skin texture. Strong texture similarity can hide a changed face shape.

One person in a new scene

Place the person from image 2 beside the cafe table in image 1. Match the scene lighting and perspective; preserve the person’s face and outfit.

Use the two-reference workflow: scene first, person second. Match the light direction in the source images when possible.

A restrained anime restyle

Restyle this image as a clean cel-shaded anime illustration. Preserve the composition, pose, outfit colors and main background objects.

Expect deliberate texture changes. Judge composition and character design separately from photorealistic likeness.

Remove a distracting object

Remove the orange traffic cone at the lower right. Reconstruct the pavement naturally and keep the person, storefront and camera position unchanged.

Switch to Raw with real guidance; do not keep raising Turbo steps when the object refuses to disappear.

Expand the canvas

Extend this scene horizontally with more of the same cafe interior on both sides. Keep a single copy of the person and preserve the central composition.

Set a wider target within the pixel budget, use fit and Euler, then inspect seams and duplicated subjects.

Try a reference garment

Dress the person in the jacket shown in the second reference. Preserve the face, pose, trousers and background; match the jacket’s cut and material.

Treat this as visual try-on, not a guarantee of fit. Check cuffs, fasteners and patterns before using it in a catalog.

Build a character reference sheet

Create a three-view reference sheet of this fictional character: front, side and back, on a plain light gray background. Keep the same outfit, colors and proportions.

This is a harder compound edit. Check every view for invented accessories or inconsistent anatomy.

Separate compound jobs into reviewable passes—recolor, then background, then upscale—but keep approved intermediate files. Chaining also accumulates errors. If the face or product drifts, restart from the last acceptable source instead of trying to repair a degraded result indefinitely.

Online access, API, pricing and GPU requirements

Online: verify the exact model name

Sogni’s model page explicitly lists Identity Edit v1.2. In Sogni Web, select that model, provide the source/context images and enter the edit instruction. The standard deployment ID is krea2_identity_edit_v1_2; Dark Beast and Sogni-specific variants are separate choices.

Pricing sources: model-specific pricing / current plans.

For integration, follow Sogni’s developer tutorials to the Python Krea Identity Edit example and current authentication instructions. A generic Krea 2 text-to-image endpoint is not interchangeable with this edit workflow. Verify one job’s estimate and result before batching.

GPU and speed: distinguish file size from working memory

Sogni’s worker guide lists 16GB VRAM for its Identity Edit deployment. That describes its optimized configuration, not a universal ComfyUI minimum. Base precision, encoder, resolution, references and offloading all matter. On an existing 16GB setup, FP8 at about 1MP and batch 1 is a sensible evaluation starting point—not a hardware guarantee.

The README gives an approximate one-minute Turbo example at 2MP, without a standardized GPU comparison. Separate cold loading from warm runs and time the same source, seed and settings. Local use has no monthly generation quota; memory, throughput and electricity still set practical limits.

Quality and real community feedback

A July 22 Reddit before/after thread praises short prompts for expression, prop and clothing changes. An August 11 user report describes distorted faces, invented clothing patterns and harsh scene boundaries. These experiences support trying it for bounded edits, not treating a good showcase as a universal likeness guarantee.

An August 30 report says full-body references kept pulling the result toward the original pose, while a face crop allowed more change. It used a customized stack, so treat it as a useful troubleshooting hypothesis rather than a controlled benchmark.

The maintainer also acknowledges distinctive facial proportions, multiple people, clothing compliance and changes outside the intended region as limitations. Inspect small text and logos manually. This review does not assign benchmark scores or claim a measured Japanese-text editing advantage.

See Sogni’s published tutorial and editing examples

Krea 2 Edit vs Qwen, FLUX and Krea Edit

Qwen-Image-Edit-2511

Qwen’s official model card describes consistency and drift improvements. Compare several recolor, removal and placement attempts using the same inputs, counting failures as well as the best output.

FLUX.2 [klein]

Black Forest Labs’ model page separates variant specifications. Its hardware-specific speed figures should not be compared directly with a differently configured Krea workflow.

Krea Web Edit

Krea’s official image editor is a hosted service. Keep Krea 2 generation, Krea’s Edit product and the community Identity Edit LoRA separate when comparing features or subscriptions.

Troubleshooting in a useful order

For two people, the model card suggests chained inserts while the node README recommends simultaneous placement. The guidance is inconsistent. Test both for a critical two-person job instead of promising that either method always works.

NSFW support and commercial-use conditions

The Identity Edit author explicitly says it is trained on SFW data and does not support NSFW concepts. Local execution is not evidence of an NSFW-specialist model. Sogni’s “Uncensored” label and Dark Beast option describe provider variants/settings, not a change to the upstream author’s stated training scope.

The node code is Apache 2.0; the weights use the Krea 2 Community License. The model card describes a commercial revenue threshold, currently below US$1 million annually, plus other obligations. Do not infer unrestricted commercial deployment from the code license. Read the weights license and your provider’s terms.

FAQ and the first experiment to run

Can I download only the LoRA?

No. You also need the Krea base, Qwen3-VL encoder, VAE and custom nodes. Select your real filenames in the supplied graph.

Should I write an ordinary negative prompt?

At Turbo CFG 1, increasing ordinary negative wording is not the useful fix. For Raw with CFG above 1, begin with the author’s image-grounded empty-prompt negative.

Is it better for anime or real photos?

The base and stacked LoRAs affect the look. Photographic restaging examples are plentiful, but they do not establish an anime benchmark. Test your own character design, outlines and clothing.

Can it reliably correct text inside an image?

This review does not establish a text-editing accuracy score. Proofread even short words; exact copy and logos may be better composited with a conventional editor.

What should an existing Krea user try first?

Use one source for three separate tasks: recolor, lighting/expression, and removal. Record edit success, likeness, background drift and time, with Turbo and Raw in their appropriate roles.

Krea 2 Identity Edit is a useful extension for taking a liked generation into the next production step. Find a repeatable one-image, one-change use case and save that preset before expanding the workflow. Keep inspection and conventional editing in the pipeline whenever exact identity or product details matter.

Related: our Krea 2 overview.

Review scope: model documentation, the supplied JSON, node implementation notes, provider information and public user experiences. No original GPU inference benchmark was run for this article. Prompts are suggested applications, not guaranteed outputs. The cover is an AI-generated editorial illustration.