Learning Representations to Guide Pixel Diffusion Transformers
The paper argues that stronger visual representations are not automatically better guides for image generation.
The authors test pixel diffusion transformers and find that externally aligned semantics can fade during pixel synthesis. Their proposed Representation Guidance method uses a low-dimensional bottleneck and feeds the resulting internal representation back into decoder generation. They report better semantic and spatial structure while keeping a plain Transformer design. On benchmarks, RG-XL/16 reaches 1.55 FID on ImageNet 256x256 and extends to text-to-image with 0.89 on GenEval and 87.1 on DPG-Bench. HF Daily Papers' note
The authors test pixel diffusion transformers and find that externally aligned semantics can fade during pixel synthesis. Their proposed Representation Guidance method uses a low-dimensional bottleneck and feeds the resulting internal representation back into decoder generation. They report better semantic and spatial structure while keeping a plain Transformer design. On benchmarks, RG-XL/16 reaches 1.55 FID on ImageNet 256x256 and extends to text-to-image with 0.89 on GenEval and 87.1 on DPG-Bench. HF Daily Papers' note
score 4