🇲🇽 Spanish, 18 October 2026, image generation
New AI Method Overcomes Visual Bottleneck in Direct Image Generation
A new adaptive framework optimizes direct pixel-level image generation models without computational overhead.
Investigadores en computación visual have presented a novel technical framework called MFlow, accompanied by the ConvRefiner module, designed to transform the way generative models create images directly at the pixel level.
The goal of this advance is to eliminate the cumbersome intermediate compressors used by current systems, achieving a direct visual synthesis, more faithful and without geometric fidelity losses in both static images and video.
The technique MFlow acts as a conductor that does not impose a fixed score: instead of randomly choosing between predicting the final clean image or calculating the speed of change, it learns to adaptively interpolate between both according to the direction and complexity of visual data.