New Research Could Fix Weird AI Images
Rice University researchers have developed a new method, ElasticDiffusion, to improve image generation in AI models, addressing common issues like distortion and repetitive details in non-square images. Current diffusion models, such as Stable Diffusion and DALL-E, struggle with aspect ratios other than square, often resulting in anomalies like extra fingers or stretched objects.
ElasticDiffusion, created by doctoral student Moayed Haji Ali, separates the local and global information in images, allowing the model to better handle different aspect ratios without retraining. By processing image details and overall composition separately, the method avoids the confusion that causes repetitive or distorted elements in AI-generated images.
However, ElasticDiffusion takes longer to generate images compared to existing models, with a goal to eventually match their speed. This approach could ultimately allow AI models to produce cleaner images across varied aspect ratios, without additional training costs.