Boosting Generative Image Modeling via Joint Image-Feature Synthesis
This work addresses a key bottleneck in generative image modeling for AI researchers, offering a novel approach that simplifies training and enhances performance, though it is incremental in building upon existing diffusion models.
The paper tackles the challenge of integrating representation learning with generative modeling in latent diffusion models by introducing a framework that jointly models image latents and semantic features, resulting in improved image quality and faster training convergence.
Latent diffusion models (LDMs) dominate high-quality image generation, yet integrating representation learning with generative modeling remains a challenge. We introduce a novel generative image modeling framework that seamlessly bridges this gap by leveraging a diffusion model to jointly model low-level image latents (from a variational autoencoder) and high-level semantic features (from a pretrained self-supervised encoder like DINO). Our latent-semantic diffusion approach learns to generate coherent image-feature pairs from pure noise, significantly enhancing both generative quality and training efficiency, all while requiring only minimal modifications to standard Diffusion Transformer architectures. By eliminating the need for complex distillation objectives, our unified design simplifies training and unlocks a powerful new inference strategy: Representation Guidance, which leverages learned semantics to steer and refine image generation. Evaluated in both conditional and unconditional settings, our method delivers substantial improvements in image quality and training convergence speed, establishing a new direction for representation-aware generative modeling. Project page and code: https://representationdiffusion.github.io