AI image generation skips compression, yet retouching results don't improve.

Beating AI News Flash: Spanish AI startup Speridlabs tried a new approach to image generation. Many AI image generation models today first generate visuals in compressed image information and then restore them into images. This process may lose details. Speridlabs wanted to know: if this step is skipped and every pixel is generated directly, would the results be better?

To this end, the team trained the open-source model Iris-3B and also adapted another existing model into a version that generates pixels directly. They had these models restore low-resolution images, judge the distance of objects in photos, and then compared them with the original approach. None showed the expected stable improvement, and some metrics were even slightly worse. The model that generates pixels directly also showed slight checkerboard artifacts in some test results.

However, this research also proved the feasibility of directly generating pixels. Iris-3B scored 0.540 in an image generation evaluation, roughly on par with Qwen-Image's 0.539. Simply removing the image compression step did not give it a clear advantage in tasks with higher detail requirements.