A new generation of image-generation models is rejecting the polished, homogeneous look that has come to define AI art, instead prioritizing creative exploration and stylistic diversity. Krea 2, a series of foundation models released today, represents a significant philosophical shift in how image-generation systems are built and trained. Rather than optimizing for a single "default" aesthetic, Krea 2 exposes users to a broad visual space and gives them practical tools to navigate it using both text and image-based controls.

The motivation behind this approach stems from a real problem in the field. Over the past few years, diffusion and flow-matching models have achieved remarkable technical feats: they can generate high-resolution images with sharp photorealism, stable structure, dense text rendering, and broad world knowledge. Yet as the industry has optimized for reliability on these capabilities, many systems have converged toward a narrow set of default aesthetics. While effective as production tools, this convergence makes them less useful as engines for creative exploration, where creators often need to search across styles, moods, compositions, and visual directions.

What Makes Krea 2 Different From Other Image Generators?

Krea 2 addresses this limitation through several architectural and training innovations. The model incorporates a diffusion transformer (DiT) architecture refined through extensive testing, along with several components designed to accelerate training and improve stability. These include improved VAEs (variational autoencoders, a type of neural network), grouped-query attention, sigmoid-gated attention, and lightweight timestep modulation. The model also integrates Qwen3-VL, an advanced vision-language system, to better understand visual content.

Performance-wise, Krea 2 ranks among the top 10 models on the Artificial Analysis leaderboard for text-to-image generation, and scores second place among models from independent labs. This competitive performance is notable because it was achieved while prioritizing creative flexibility rather than chasing a single polished default.

To reduce the gap between how the model was trained and how users actually express creative intent, Krea 2 includes two key systems. A prompt expander maps simple or underspecified user prompts into richer visual directions without overwriting the user's original intent. It is trained through a two-stage supervised fine-tuning (SFT) and reinforcement learning (RL) pipeline on top of open-source large language models (LLMs), with the objective of improving image quality while encouraging creative variation and controllable exploration. Complementing this textual interface, a style-reference system lets users express visual intent through images when words are insufficient, allowing them to inject the style or mood of reference images with minimal content leakage and fine-grained control over style strength and weighted style mixing.

How Does Krea 2's Training Data Differ From Competitors?

One of the most distinctive aspects of Krea 2 is its approach to data curation. Rather than relying on conventional model-based filtering that uses aesthetic scores and image-quality-assessment models, Krea 2's creators built a large-scale data infrastructure and distributed training framework from scratch. They argue that conventional filtering introduces implicit biases; for example, such methods may classify a blurry image as low quality, even though motion blur or softness can be a deliberate artistic choice.

The team filters out only specific categories of problematic data:

• Duplicated Samples: Removing duplicate images and over-represented concepts to ensure diversity in the training set.
• Poor Alignment: Excluding samples for which vision-language models consistently fail to capture important aspects of the image.
• Undesired Biases: Filtering out samples that induce undesired biases and artifacts in the model's output.
• High Complexity: Removing samples with visual complexity too difficult to model reliably at low resolution.
• Synthetic Images: Excluding AI-generated samples entirely, as the team found that even a small proportion of synthetic images introduces biases into the model's output distribution.

This last point is particularly significant. The team designed in-house classifiers to filter out AI-generated images because synthetic data, while an effective shortcut for acquiring model capabilities, tends to be easier to learn from. This effectively imposes an upper bound on model quality. By excluding synthetic images entirely, Krea 2 maintains a higher ceiling for what the model can achieve.

The captioning process also reflects a commitment to richness and diversity. The team runs an optical character recognition (OCR) model on each image to extract visible text, then provides both the OCR results and available metadata to a captioning model, which produces an enriched caption incorporating world knowledge alongside the extracted text. Once a context-rich, long-form natural-language caption is obtained, a cheaper LLM reformats it into a variety of lengths and formats, exposing the model to a range of prompt styles.

Steps to Understanding Krea 2's Multi-Stage Training Approach

Krea 2 employs a sophisticated multi-stage pipeline to progressively refine the model's output distribution:

• Pretraining Stage: The model learns basic text-image alignment and structure using billions of images at low resolution (256 pixels), building core capabilities efficiently before scaling up.
• Midtraining Stage: The model is exposed to progressively higher resolutions (512 pixels and 1024 pixels) through a curriculum-learning strategy, dedicating the majority of computing resources to low-resolution stages before equipping the model with high-fidelity generation capabilities.
• Supervised Fine-Tuning: The model is refined using carefully labeled examples to improve performance on specific tasks and user preferences.
• Preference Optimization: The model learns which outputs users prefer, allowing it to align better with human expectations.
• Reinforcement Learning: The model is further refined through reward signals that encourage desired behaviors and discourage undesired ones.

At the low-resolution pretraining stage, the team relies heavily on inexpensive CPU-based filters to remove low-quality images. These range from simple broken-file, resolution, and aspect-ratio filters to more sophisticated Laplacian filters that remove images with extreme textures and noise patterns. During development, the team encountered specific challenges, such as a tendency for the model to generate flat-color backgrounds and border artifacts. To mitigate this, they used RGB entropy and white/black pixel ratios as filtering criteria.

The shift toward exploratory generation represents a meaningful departure from how much of the industry has approached image-generation models. Rather than treating the model as a tool for reliably producing a single polished aesthetic, Krea 2 treats it as a medium for creative exploration. This philosophy extends to how the model is evaluated and deployed, with success measured not just by benchmark scores but by the breadth of visual space the model can access and the precision with which users can navigate it.

For creators and professionals who have felt constrained by the homogeneous outputs of existing image-generation tools, this approach offers a different path forward. By building a model that exposes a broad visual space and providing practical ways to move through it, Krea 2 suggests that the next frontier in generative AI may not be about achieving higher quality in a narrow sense, but about enabling richer creative expression across a wider range of aesthetics and styles.