**[AI-Generated Images Skip Compression, But Editing Results Aren't Better]** Beating AI News: Spanish AI startup Speridlabs has experimented with a new approach to generating images. Currently, many AI image generation models first create visuals within compressed image data and then restore them into full images, a process that may result in detail loss. Speridlabs wanted to find out if skipping this step and directly generating each pixel would yield better results. To explore this, the team trained an open-source model called Iris-3B and modified another existing model into a version that directly generates pixels. They tested these models on tasks like enhancing low-resolution images and determining the depth of objects in photos, comparing the results with traditional methods. However, the outcomes did not show the expected consistent improvements, and some metrics were slightly worse. In certain test results, the pixel-direct models even exhibited minor grid artifacts. That said, the research demonstrated the feasibility of directly generating pixels. In one image generation evaluation, Iris-3B scored 0.540, nearly matching Qwen-Image's 0.539. However, removing the image compression step did not provide a significant advantage in tasks requiring higher detail precision. [Original Article Link]