Coevolving Representation Diffusion — adaptive semantic space improves FID by 12% on ImageNet 256x256

CoReDi (Coevolving Representation Diffusion) lets the semantic representation space adapt during diffusion training via a lightweight linear projection, improving FID by 12% on ImageNet 256x256 compared to fixed-space baselines. Stable coevolution is achieved through stop-gradient, normalization, and regularization to prevent collapse. The method works for both VAE latent and pixel-space diffusion, but requires a pre-trained visual encoder and careful tuning of regularization hyperparameters.

• #diffusion
• #representation-learning
• #image-generation
• #imagenet