Krea releases Krea 3 image-editing model
Krea has released Krea 3, an image-to-image model it says matches frontier rivals on quality while finishing edits within eight seconds, roughly three to five times faster than Nano Banana Pro.
Krea detailed its new Krea 3 model in a technical report. Krea 3 is a foundation image-to-image (I2I) model built to edit images based on reference inputs. The company says it keeps characters and products consistent across edits, changes only the targeted parts of an image, and can combine up to eight reference images in one generation.
According to Krea, the Krea 3 Ultra variant debuted at #4 on all independent arenas, which the company says makes it the #2 image generation lab. The report says production edits finish within 8 seconds. Krea says Nano Banana Pro takes about 20 to 40 seconds for a comparable edit in its own production traffic, and that GPT Image models can take up to several minutes. Krea also says the model is competitive with Nano Banana Pro and GPT Image 2 at "a fraction of the price," but the report excerpt gives no specific pricing. The speed and cost comparisons are Krea's own measurements.
Krea calls the architecture "deliberately standard" and says the main work is in training, data and post-training. Krea 3 starts from the Krea 2 pretrain and continues pretraining on a larger text-to-image dataset that includes native 2K samples. Reference images are VAE-encoded and appended to the diffusion transformer's image sequence as clean tokens with timestep set to zero. Krea says this allows key–value caching after distillation. A vision-language model text encoder reads the reference images alongside the edit instruction, and only the instruction tokens go to the DiT.
On the training side, the report describes sampling reference and target resolutions separately and using length-binned sequence packing to handle varying reference counts. The available excerpt ends before the post-training details.