PixelDense improves pixel diffusion by separating semantic and geometric features
PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion
Computer Vision and Pattern Recognition
Summary
Diffusion models that generate images learn faster and create better results when trained using guidance from other models. This paper finds that using models which focus on spatial structure and geometry, rather than just overall meaning, improves training. The researchers introduce PixelDense, a method that separates semantic and geometric information into different streams to prevent conflicts during training. This approach leads to faster learning and improved image editing and reconstruction quality.
What this means in practice
- •For computer vision engineers: Improve the training speed and accuracy of image generation models by separately processing semantic and geometric features.
- •For image editing software developers: Create more precise image editing tools that maintain background and layout better during edits using the PixelDense approach.
Authors
Lehan Yang, Daiqing Qi, Wenhao Zhang, Avery Li, Yiqing Yang, Yifan Li, Yu Kong, Haitian Zheng, Zhifei Zhang, Zhe Lin, Varun Jampani, Sheng Li
Abstract
Representation alignment (REPA) accelerates diffusion transformer training, but its alignment targets are almost exclusively semantic encoders such as DINOv2 and CLIP. Recent analysis points to spatial structure, not global semantics, as the carrier of the alignment effect, yet dense-prediction foundation models trained to predict that structure remain overlooked as REPA targets. In pixel-space diffusion, SAM2, Depth Anything v2, and Metric3D v2 each outperform the DINOv2-only GenEval baseline, with the two geometric teachers leading the segmentation teacher. A flat sum of all four teachers, however, lands below the best single geometric teacher, as semantic and geometric gradients compete for one denoiser projection. We introduce PixelDense, which routes DINOv2 and SAM2 through a semantic projection stream, routes Depth Anything v2 and Metric3D v2 through a geometric projection stream, and adds a weight-space orthogonality penalty that keeps the two streams in disjoint subspaces. All four teachers are frozen during training and dropped at inference. Applied to PixelGen and DeCo with a single recipe, PixelDense improves GenEval, DPG-Bench, and HPS v2.1, raises PixelGen-XXL's GenEval Overall from 0.7927 to 0.8093, and beats every single-teacher and unfactored multi-teacher variant. In partial-noise reconstruction, independent panoptic, depth, and surface-normal probes show up to 53.1% PQ gain and 36.0% depth AbsRel reduction at $τ=0.5$ across COCO and Flickr30K. From random initialization, PixelDense also reaches the baseline's peak GenEval 1.23x faster. In SDEdit editing on PIE-Bench, PixelDense keeps more of the source background and layout at every edit strength, raising background PSNR by up to 2.2 dB.