Research map / Image generation / Autoregressive and latent generation
Scaling Dense Prediction with Latent Decoding
Xiaoyang Wu, Yixing Lao, Chengyao Wang, Senqiao Yang, Yujia Zhang, Hengshuang ZHAO
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed-schedule fusion with stage-wise adaptive selection of encoder features (ablation) | L-DPT · Hypersim val relative depth | ARel: 9.67 -> 8.89 | Improved |
| use stage-specific selectors instead of one shared selector (ablation) | L-DPT · Hypersim val relative depth | delta1: 92.06 -> 92.76 | Improved |
| replace DPT decoder with L-DPT S (23M vs 230M) on frozen DINOv3 ViT-7B | DINOv3 depther · DIODE zero-shot | ARel: 25.6 -> 22.6 | Improved |
| replace DPT decoder with L-DPT S on frozen DINOv3 ViT-7B | DINOv3 depther · NYUv2 zero-shot | ARel: 4.3 -> 5.0 | Hurt |
| scale L-DPT decoder from S to L (23M to 269M) (ablation) | L-DPT · DIODE zero-shot | ARel: 22.6 -> 22.0 | Improved |
| replace DINOv3 ViT-Adapter + Mask2Former with L-DPT S decoder | DINOv3 ViT-7B segmentor · PASCAL VOC 2012 (simple) | mIoU: 90.1 -> 93.4 | Improved |
| replace DINOv3 segmentor with L-DPT L on frozen ViT-7B | DINOv3 ViT-7B segmentor · COCO-Stuff 164k (simple) | mIoU: 53.8 -> 54.0 | Improved |
| use latent decoding instead of DPT decoder at high resolution | dense decoder · decoder peak memory at high end of resolution sweep | peak memory (GB): 59.4 -> 15.3 | Improved |
| use latent decoding instead of DPT decoder at high resolution | dense decoder · decoder latency at high end of resolution sweep | latency (s): 394 -> 5.3 | Improved |
| use L-DPT-L instead of InfiniDepth with DINOv3-L | relative depth · ETH3D delta1 | delta1: 98.6 -> 97.4 | Hurt |
Closest papers by meaning
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.87)
- L2P: Unlocking Latent Potential for Pixel Generation (NeurIPS 2026 · similarity 0.84)
- Efficient Multitask Dense Predictor via Binarization (CVPR 2024 · similarity 0.84)
- ZPressor: Bottleneck-Aware Compression for Scalable Feed-Forward 3DGS (NeurIPS 2025 · similarity 0.84)
- DenseMLLM: Standard Multimodal LLMs for Dense Prediction (ICML 2026 · similarity 0.84)
- TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
