AI-generated images skip compression, but the retouching effect is no better.

Beating AI News Insight: Spanish AI startup Speridlabs has tested a novel AI image generation method. Most current AI image models first generate content from compressed image data before reconstructing the final image, a process that can cause detail loss. Speridlabs sought to determine if skipping this step to generate pixels directly would deliver better results. The team trained an open-source model, Iris-3B, and adapted another existing model to generate pixels directly. They evaluated these models on tasks including low-resolution image restoration and judging object distance in photos, comparing their performance with existing approaches. The results did not show the expected consistent improvement; some metrics were even slightly worse. The direct pixel-generating models also exhibited minor grid artifacts in some test outputs. However, the research confirmed the feasibility of direct pixel generation. Iris-3B scored 0.540 in an image generation benchmark, nearly matching Qwen-Image’s 0.539. Simply removing the image compression step did not give it a clear edge in detail-intensive tasks.