Krea publishes Krea 2 technical report

Krea has published a technical report on Krea 2, a series of image-generation foundation models. The company argues that many current systems have "converged toward a narrow set of default aesthetics," which makes them weaker tools for creative exploration.

Krea built its data infrastructure and distributed training framework from scratch. Training runs in stages: pretraining, midtraining, supervised finetuning, preference optimization and reinforcement learning. The model is a diffusion transformer that uses iREPA, improved VAEs and Qwen3-VL to speed up convergence. The report says grouped-query attention, sigmoid-gated attention, lightweight timestep modulation and multilayer aggregation of text-encoder features improved training stability and efficiency.

The data strategy is the report's most distinctive claim. Krea says aesthetic-score and image-quality filters bring in hidden biases. Such filters might reject motion blur, for example, even when the blur is a deliberate artistic choice. Krea therefore removes only duplicates and over-represented concepts, poorly captioned samples, images that cause bias or artifacts, images too complex to learn at low resolution, and AI-generated images. Krea says even a small share of synthetic images biases the model's outputs and "effectively imposes an upper bound on model quality."

To connect short user prompts with the dense captions used in training, Krea built two more systems. One is a prompt expander, trained with SFT and RL on top of open-source LLMs. The other is a style-reference system that supports adjustable style strength and weighted mixing of multiple reference images.

Krea says Krea 2 ranks in the top 10 of the Artificial Analysis text-to-image leaderboard and second among models from independent labs. The available excerpt ends partway through the data section, so later details of the report are not covered here.