PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion

NeurIPS 2026

📄 Paper | 🌐 Project Page | 💻 Code | 🤗 Model

PixelDense aligns a pixel-space diffusion transformer to frozen dense-prediction models during training: DINOv2 and SAM2 supervise a semantic projection stream, Depth Anything v2 and Metric3D v2 supervise a geometric projection stream, and an orthogonality penalty keeps the two streams apart. The teachers are used only during training, so this demo runs the plain 1.1B pixel-space denoiser (PixelGen-XXL architecture) with the Qwen3-1.7B text encoder and the 25-step Adam LM sampler, generating 512 × 512 images directly in RGB space.

Example prompts

@inproceedings{yang2026pixeldense,
  title={PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion},
  author={Yang, Lehan and Qi, Daiqing and Zhang, Wenhao and Li, Avery and Yang, Yiqing and Li, Yifan and Kong, Yu and Zheng, Haitian and Zhang, Zhifei and Lin, Zhe and Jampani, Varun and Li, Sheng},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}