Apple released an open-source AI model called DiffuCode-7B-cpGRPO that uses a diffusion-based approach for code generation, allowing faster and more globally coherent code output by generating out of order and refining multiple chunks simultaneously, building on Alibaba's foundation model and recent research in diffusion architectures.
Researchers have discovered that the apparent creativity of diffusion models in AI image generation stems from deterministic imperfections in their denoising process, revealing that their 'creativity' is an inevitable outcome of their architecture, similar to biological morphogenesis processes. This insight could impact future AI development and our understanding of human creativity.
Google researchers proposed a method that combines a standard autoencoder with a diffusion process to efficiently compress high-quality images using score-based generative models. The proposed method outperforms several recent generative approaches in terms of image quality and preserves details much better compared to state-of-the-art approaches. The study revealed specific details that can be useful for future research in this domain, such as the impact of noise schedule and the amount of noise injected during image generation.