Krea 2: the open image model you run yourself

Krea 2 ships as two models, and picking the wrong one wastes time. K2 Raw is the undistilled base: Krea describes it as being for fine-tuning and post-training research. It is what you train LoRAs against. K2 Turbo is guidance- and timestep-distilled for fast few-step generation, and it is what you should generate with day to day.

Krea's own recommended workflow ties them together: train your LoRAs on Raw, then apply those LoRAs to Turbo for fast inference. Krea publishes its own trained LoRAs alongside the models, and they work on both. The community has added more since.

In practice Turbo produces a usable image in about 8 steps. Benji's AI Playground, who tested it locally at release, generates at 2K resolution in a single pass on Turbo and describes it as fast and smooth without the GPU straining. No upscaler stage in the loop.

This is the capability worth knowing about. Hands are the classic failure mode of open image models, the thing that marks a generation as AI at a glance. Benji's hands-on assessment is that Krea 2 renders hands and human anatomy better than Flux, which he says regularly mangles character hands, and he notes people already using Krea 2 for influencer-style portrait work.

Treat that as one reviewer's opinion rather than a benchmark, because that is what it is. But it points at why the leaderboard position translates into something useful: a model that wins on paper and still gives your subject six fingers is not a model you can build on.

The style range is genuinely broad, too. Krea's launch material leads with a single cowboy prompt rendered across wildly different graphical treatments, and independent testing backs it up.

Per Krea's technical report: text rendering and dense visual detail, a prompt expander that maps short input into richer captions, and the style-reference system above. It was trained at 256px, 512px and 1024px.

Out of the box it is text-to-image only, with no editing. The community filled that gap: an unofficial fine-tune on Hugging Face called Krea 2 Identity Edit (v1 and v1.1) turns it into an instruction-based, identity-preserving editor driven by plain English. Swap a background, change an outfit as a virtual try-on, move a person into a new scene. It runs on both Raw and Turbo and needs a third-party ComfyUI node pack installed by hand from GitHub. Worth being clear that this is a community fine-tune, not a Krea product: Krea's support and technical report do not cover it.

Krea 2 gets described as uncensored across YouTube and the forums, so it is worth being precise about what is actually true.

Krea's technical report makes no uncensored claim. It does not discuss safety alignment or content filtering at all, in either direction. The framing traces to a structural fact rather than a feature: open weights running on your own hardware have no hosted content filter, because there is no server in the loop to enforce one. That is a property of running any open model locally, not something Krea built or advertised.

The related fact worth stating plainly is that the community trains its own LoRAs and fine-tunes against Krea 2, and Identity Edit is one public example of that pipeline working. A LoRA can teach a base model material it was not trained on. That is how open-weight ecosystems work in general, and it is not a Krea 2 capability.

If you have a modern GPU and want image quality in the same conversation as the paid APIs, without a subscription and without sending prompts to anyone's server, this is the most interesting open release of the period. The FP8 and NVFP quantised builds that ComfyUI merged are what bring it within reach of consumer hardware.

If you do not have the GPU or the patience for a ComfyUI setup, Krea runs the model on its own hosted platform with a free signup, which is a lower-friction way to see whether the quality justifies the install.