Research map / Image generation / Autoregressive and latent generation
Pixel-Level Residual Diffusion Transformer: Scalable 3D CT Volume Generation
Zhenkai Zhang, Markus Hiller, Krista A. Ehinger, Tom Drummond
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 |
|---|---|---|---|
| remove patch overlap (add overlap) (ablation) | PRDiT-4L · LIDC-IDRI | 3D FID (x10^3): 3.27 -> 2.04 | Improved |
| add local denoiser (ablation) | PRDiT-4L · LIDC-IDRI | 3D FID (x10^3): 3.10 -> 2.04 | Improved |
| add global residual DiT to local component (ablation) | PRDiT-4L · LIDC-IDRI | 3D FID (x10^3): 41.92 -> 2.04 | Improved |
| use hot predictor step k=2 instead of cold k=1 (ablation) | PRDiT-4L · LIDC-IDRI | 3D FID (x10^3): 8.889 -> 2.173 | Improved |
| increase predictor step from k=2 to k=4 (ablation) | PRDiT-4L · LIDC-IDRI | 3D FID (x10^3): 2.173 -> 4.184 | Hurt |
| increase depth from 4 to 12 layers | PRDiT · LIDC-IDRI 128^3 | 3D FID (x10^3): 2.04 -> 1.41 | Improved |
| increase depth from 4 to 12 layers | PRDiT · RAD-ChestCT 128^3 | 3D FID (x10^3): 1.92 -> 1.45 | Improved |
| train at 64^3 then upsample to 128^3 instead of training from scratch (ablation) | PRDiT · LIDC-IDRI | 3D FID (x10^3): 2.04 -> 2.89 | Hurt |
| train at 64^3 then upsample to 128^3 instead of training from scratch (ablation) | PRDiT · LIDC-IDRI | training cost: 80 GPUh -> 12 GPUh | Improved |
| fine-tune full backbone instead of freezing it at 256^3 (ablation) | PRDiT · LIDC-IDRI 256^3 | 3D FID (x10^3): 2.280 -> 1.702 | Improved |
| hot predictor-corrector k=3.5 instead of conventional hot diffusion (ablation) | PRDiT · LIDC-IDRI | 3D FID (x10^3): 6.456 -> 5.718 | Improved |
| use PRDiT-4L instead of baselines | PRDiT-4L · LIDC-IDRI 128^3 | 3D FID (x10^3): 3.26 -> 2.04 | Improved |
Closest papers by meaning
- DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction (NeurIPS 2024 · similarity 0.87)
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.86)
- Scaling Diffusion Mamba with Bidirectional SSMs for Efficient 3D Shape Generation (AAAI 2025 · similarity 0.86)
- TCAM-Diff: Triplane-Aware Cross-Attention Medical Diffusion Model (AAAI 2025 · similarity 0.86)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.86)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.85)
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.
