Research map / Image generation / Autoregressive and latent generation
PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation
Haofei Xu, Rundi Wu, Philipp Henzler, Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Marc Pollefeys, Andreas Geiger, Federico Tombari, Michael Niemeyer
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 |
|---|---|---|---|
| predict clean x instead of velocity v (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | Rel depth: 24.07 -> 5.54 | Improved |
| add random zero timestep with 10% probability to noise schedule (t-shift -0.8) (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | Rel depth: 7.80 -> 6.00 | Improved |
| use DINOv3 last layer instead of linear patch embedding (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | Rel depth: 9.64 -> 6.00 | Improved |
| use DINOv3 instead of DINOv2 last-layer features (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | BF1: 5.11 -> 7.24 | Improved |
| sample 4 intermediate DINOv3 layers instead of last layer (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | BF1: 7.24 -> 13.47 | Improved |
| use MoGe-2 4-layer features instead of DINOv3 (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | delta1 point: 91.18 -> 93.35 | Improved |
| use MoGe-2 4-layer features instead of DINOv3 (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | BF1: 13.47 -> 11.75 | Hurt |
| add relative point loss to v-loss (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | BF1: 13.47 -> 13.92 | Improved |
| add relative point loss to v-loss (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | delta1 depth: 95.08 -> 94.88 | Hurt |
| use generative flow matching instead of deterministic regression (ablation) | PointDiT-L · Ablation 256x256 avg 7 sets | BF1: 10.90 -> 13.92 | Improved |
| use patch size 16 instead of 32 at 512x512 (ablation) | PointDiT-L · 512x512 6-dataset subset | BF1: 6.17 -> 10.37 | Improved |
| use patch size 16 instead of 32 at 512x512 (ablation) | PointDiT-L · 512x512 6-dataset subset | Rel depth: 3.48 -> 3.06 | Improved |
Closest papers by meaning
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.89)
- Coherent 3D Scene Diffusion From a Single RGB Image (NeurIPS 2024 · similarity 0.89)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.89)
- Pixel-Perfect Depth with Semantics-Prompted Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- SurGe: Improved Surface Geometry in Point Maps (NeurIPS 2026 · similarity 0.88)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.88)
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.
