Jan Sverre Bauge

Running Krea 2 Identity Edit locally – image quality close to top cloud services

I ran 20 local Krea 2 Identity Edit generations and finished 16 scene ideas. The image quality impressed me, while four misses exposed the model's practical limits.

Jan Sverre shown across multiple Krea 2 Identity Edit scenes generated locally from one reference portrait
Krea 2 Identity Edit turned one portrait into dramatically different scenes while preserving a recognizable reference appearance.

Krea 2 Identity Edit is the first local image-to-image solution I have tested that genuinely made me think the gap to the best cloud services is getting small. I gave the model one square portrait and asked it to move the same person into everything from a Norwegian fjord and a Western town to a motorcycle race, the Moon and a kitchen full of tiny robots.

In total, I ran 20 documented test generations. They cover 16 finished scene ideas: four in the first test and 12 in the expanded set. Four jobs in the expanded set were misses that had to be replaced. That matters, because this is not a model that follows every instruction perfectly on the first attempt.

Even so, the overall impression is unusually strong. The model kept the person recognizable through action, profile views, dramatic lighting, new clothes, full-body compositions and completely absurd scenes. My subjective quality judgment remains that several images are up there with Nano Banana Pro. This was not a controlled A/B test or a benchmark. It was my visual assessment of the results I actually produced locally.

Local image-to-image that survives major scene changes

Krea 2 Identity Edit v1.2 is an unofficial community LoRA that turns Krea 2 into an instruction-driven image editor. You provide a reference image and a plain-text instruction, and the model tries to make the change without losing the person along the way. The model card on Hugging Face lists outpainting, inpainting, virtual outfit changes and support for different aspect ratios.

What stood out in my test was how much could change at once. A head-and-shoulders portrait became new bodies, poses, clothes, environments and image formats. The reference appearance held in three-quarter profile, under intense pink light and in action scenes where the face occupied only a small part of the frame. Look at the images before we get into the technical details:

Jan Sverre riding a dinosaur through a lush prehistoric valley
Jan Sverre riding a motorcycle through rain and neon light in a city at night
Jan Sverre looking toward a pink sunset over the Norwegian coast
Jan Sverre in modern clothes walking through a Norwegian town in 1890
Jan Sverre painting colorful street art on a legal art wall
Jan Sverre drinking morning coffee in a robe and pink slippers on the Moon
Jan Sverre sailing through a Norwegian fjord on a giant rubber duck
Jan Sverre as a pirate captain at the ship's wheel in dramatic storm light
Jan Sverre standing as a cowboy beside a horse in a Western town

The main gallery shows the pattern better than 16 nearly identical image descriptions would. The model handled profile views, speed, animals, historical settings and absurd scale surprisingly well. It also managed simple hand and object interactions. At the same time, the black hood from the reference garment leaks into the pirate and cowboy images. Strong reference conditioning is both the strength and the weakness here.

Four misses showed what the model actually struggles with

The expanded set consisted of 16 API jobs: 12 intended final scenes and four results that were replaced. Every job completed technically, but technically complete is not the same as correct. The robots lost a race they were supposed to win, the hoodie survived an explicit outfit change, and the robot chef grew too many arms twice.

The useful part was that the three types of failure required different solutions. The spatial misunderstanding was fixed with a strict composition. The outfit change needed slightly weaker reference conditioning and a more precise wardrobe description. The anatomy problem only disappeared when I stopped asking the model to perform a complicated human action.

The robot race: left and right beat “the robots are leading”

In the first version, the robots were supposed to be several meters ahead and winning. The image looked good, but the person was placed in front of the robots. In the second version, the prompt described a side view with the robots ahead on the right and the person behind on the left. That explicit layout fixed the spatial misunderstanding.

Jan Sverre running ahead of the robots in the incorrect first version of the robot race V1 – wrong winner
Jan Sverre losing the robot race as the robots cross the finish line ahead of him V2 – correct layout

Fashion: the reference garment refused to let go

Fashion V1 was a good-looking portrait, but it was not the edit I requested. The model kept the black hoodie instead of creating the open linen shirt. For V2, I lowered ref_boost from 4.0 to 3.25 and grounding_px from 768 to 704. I also described the intended outfit change more precisely. That gave the model more freedom to change the garment without serious loss of resemblance.

Jan Sverre still wearing the black hoodie in the failed first fashion version on a rooftop terrace V1 – hoodie remained
Jan Sverre in an open white linen shirt under warm sunlight on a rooftop terrace V2 – outfit change worked

The robot chef: fewer tasks beat more forbidden words

Robot Chef V1 gave me an extra arm or hand in a busy cooking scene. V2 explicitly requested “exactly two arms and exactly two hands” and assigned one task to each hand. It did not help. The anatomy only came out clean in V3, when the person stood with folded arms and the robots did the cooking. Sometimes a simpler pose is better than an ever-longer negative prompt.

Jan Sverre cooking with tiny robots but shown with an extra arm V1 – extra arm
Jan Sverre making pasta with tiny robots but still shown with an extra hand V2 – extra hand remains
Jan Sverre watching tiny robot chefs with his arms folded in a restaurant kitchen V3 – simpler pose worked

How the workflow understands both the person and the instruction

The technical core consists of a Krea 2 Turbo model, Qwen3-VL 4B, Qwen Image VAE and the Identity Edit v1.2 LoRA. I applied the LoRA at a strength of 1.0. Krea2EditModelPatch then injected the reference image as VAE-based in-context tokens, while Krea2EditGroundedEncode let Qwen3-VL interpret the instruction together with the image itself.

That separates this workflow from the regular local Krea 2 img2img branch I tried first. The old branch produced little more than a small pink sliver beneath the original garment and shifted the face toward more generic features. The Identity Edit workflow had both a stronger grip on the reference appearance and a better understanding of what actually needed to change. The ComfyUI-Krea2Edit node package provides the two specialized nodes.

Reference image passing through both VAE and Qwen3-VL before Krea 2 editing
The reference image is used both as VAE-based in-context tokens and in Qwen3-VL’s grounded instruction path before Krea 2 creates the edited image.

The default settings in the expanded set were ref_boost 4, grounding_px 768, fit, 12 steps and CFG 1. The model card recommends roughly 4 as a starting point when strong facial and body resemblance is the goal. The fashion retry also showed why these values should not be treated as sacred: slightly less reference control may be necessary when a garment genuinely needs to disappear.

The fit geometry is particularly useful when the source and output use different aspect ratios. In the first tests, I used a square 2048 x 2048 pixel portrait for both 768 x 1344 and 1344 x 768 outputs. The expanded set also used 1536 x 864, 864 x 1536 and 1024 x 1024. The model extended the scenes instead of stretching or duplicating the reference image.

About 23 seconds per image on an RTX 4090

The 20 test jobs used a total of 468.014 seconds of pure execution time. That gives an average of 23.401 seconds per job. The fastest took 13.974 seconds and the slowest 30.288 seconds. In plain English: about 23 seconds per image, with a measured range of roughly 14-30 seconds on my RTX 4090.

The numbers come from the ComfyUI history between execution_start and execution_success. Queue time is excluded. That distinction matters when several jobs run at once: automation must track the correct prompt_id and read the corresponding history entry. Pulling the “newest file” from the output directory can return the wrong image.

This is still not a universal performance benchmark. The measurements come from one machine, one local installation and one specific workflow. I did not measure peak VRAM use either, so I will not invent a minimum requirement. What I can document is that full v1.2 worked on an RTX 4090 and produced square, portrait and landscape images within this time range.

The quality is closing in on closed cloud services

I had previously tested Qwen Image Edit 2511 for local image editing. It was the best local image editor I had tested at the time, but the results were still only so-so compared with what I wanted. I had also tested FLUX.2 locally without getting usable results. Krea 2 Identity Edit moves that practical bar quite a distance.

The most convincing result is that the model could make large, simultaneous changes and still deliver images where the person remained recognizable. Action scenes, profiles, full-body views, historical clothing, animals and entirely new environments all worked. At the same time, the four misses show that a strong reference appearance does not automatically mean perfect instruction following or anatomy.

That is the basis for saying that several results gave me a visual quality impression up there with Nano Banana Pro. I did not put the systems head-to-head with the same reference, prompt and blind evaluation. This is not a win, a ranking or a controlled benchmark. It is an honest quality judgment from practical use: the local solution gave me images I would previously have expected to need one of the best cloud services to obtain.

Local execution also gives me control over the workflow, model files, LoRAs, seeds and finished PNG files. When working with my own identity or reference people who have approved the use, both the source and the result can remain on my machine. I have explored the same kind of control in local AI video with my own face, but here the work happens directly in image editing.

What does the local installation require?

This is not something you open in a browser and use in ten seconds. It requires a ComfyUI instance with Krea 2 support, the Krea 2 model itself, Qwen3-VL 4B, Qwen Image VAE, the Identity Edit LoRA and the node package. The complete model chain in my setup occupied about 19.06 GiB.

Krea 2 Turbo, Qwen3-VL and the VAE were already installed in the environment. The new model file I actually had to add was therefore about 1.70 GiB. The node package required no new Python dependencies in this environment. That is the practical installation information that matters to somebody who already has parts of a Krea 2 setup.

Rank-reduced R128 and R64 variants of about 0.91 and 0.46 GB are also available upstream. I tested neither of them. I chose full v1.2 for the best possible quality on the RTX 4090, so those smaller files are not evidence that you will get the same result or that a particular amount of VRAM is sufficient.

The limitations are clearer after 20 jobs

The test basis is much better than four simple portraits, but it still covers one person and one good, sharp reference photo. I did not test multiple people in the same edit, blurry references, demanding inpainting or a broad range of faces. The results document what this model achieved in this setup. They do not prove that it will always do the same.

The failures are quite specific. Spatial relationships can be reversed. The reference garment can leak into new costumes or overpower an outfit change. Complicated scenes involving hands, tools and several small actors can produce extra limbs. I did not fact-check the historical accuracy or the dinosaur science either. The images can be convincing without being documentary evidence.

The developer also says that distinctive facial geometry can drift toward more typical proportions, that two people can begin to resemble each other and that outfit changes do not work with every seed. Above roughly 2 megapixels, source elements can leak or be duplicated. That is worth knowing before increasing the resolution simply because the GPU allows it.

The node package uses the Apache-2.0 license, while the model weights are covered by the Krea 2 Community License Agreement. According to the model card, the license agreement for the model weights permits commercial use below an annual revenue threshold of US$1 million. Businesses above that threshold must contact Krea about an enterprise license. Read the agreement yourself before a commercial rollout – this is not legal advice.

Identity editing also requires common sense. I used my own reference photo. With other real people, consent should be a given before both generation and publication. The model developer says the project was trained on SFW data and rejects harmful, non-consensual or sexualized use of real identities.

This is local image-to-image I actually want to use

Krea 2 Identity Edit solved the main problem I gave it. One square portrait became 16 finished scene ideas with new clothes, environments, lighting, poses and aspect ratios. The tests covered 20 generations, averaged about 23 seconds on the RTX 4090 and showed that the reference appearance could survive even when the scene was dramatically rebuilt.

The four misses make the verdict more credible, not less. The model can misunderstand space, cling to the reference garment and make a mess of complicated anatomy. But those failures were fixed with better composition, adjusted reference strength or a simpler human pose. None of the solutions required me to abandon the scene idea itself.

For the first time in a while, I am not left feeling that a local image editor is mainly an entertaining technical project. This feels like a tool. Nano Banana Pro remains a subjective yardstick in this article, not a controlled opponent. But the fact that a local workflow invites the comparison at all says quite a lot.

Frequently asked questions

Can Krea 2 Identity Edit run entirely locally?

Yes. I ran Krea 2 Identity Edit v1.2 locally in ComfyUI with Krea 2 Turbo, Qwen3-VL 4B, Qwen Image VAE and the Identity Edit LoRA. The test used an RTX 4090. I did not measure a general minimum VRAM requirement.

How quickly does Krea 2 Identity Edit generate an image?

The 20 local test jobs averaged 23.401 seconds of pure execution time on an RTX 4090, with a range of 13.974-30.288 seconds. These measurements cover one specific machine and workflow, not a universal performance benchmark.

How much storage does Krea 2 Identity Edit need?

The complete model chain in my setup occupied about 19.06 GiB. Because Krea 2 Turbo, Qwen3-VL and the VAE were already installed, I only had to add about 1.70 GiB for full Identity Edit v1.2, plus the small node package.

Does Krea 2 Identity Edit get every image right on the first attempt?

No. Four of the 16 jobs in the expanded set were replaced because the model misunderstood a spatial relationship, kept the wrong garment or created extra limbs. Clearer layouts, weaker reference conditioning and simpler poses solved the three failure types.

Is Krea 2 Identity Edit better than Nano Banana Pro?

I do not have the evidence to claim that. Several results gave me a subjective quality impression up there with Nano Banana Pro, but I did not run a controlled A/B test with the same reference and prompt. The comparison is a visual judgment, not a benchmark.

Leave a Reply

Your email address will not be published. Required fields are marked *