Research map / Image generation / Image restoration
Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes two diffusion-based single-image indoor inverse rendering models (PDM for diversity, SDM for accuracy) that differ in training timestep count and channel-wise noise scheduling across geometry, material, and lighting, with a per-pixel lighting autoencoder.
Problem
Single-image indoor inverse rendering is ill-posed; deterministic methods give one solution while generative methods struggle to produce diverse samples and accurate predictions simultaneously, and latent diffusion models cannot handle per-pixel environment maps well.
Approach
Channel-wise noise scheduling assigns different noise schedules (tau) to geometry, material, and lighting channels so one diffusion architecture can be trained as either a diverse model (T=1000, PDM) or an accurate, switchable model (T=1 redefined as 3n+1 steps, SDM); lighting is encoded as a neural implicit feature vector in pixel space.
Key results
- On the synthetic test set, SDM reduced normal MSE to 1.502 (x10^-2) vs 1.738 for Zhu et al. and improved albedo MSE to 0.349 vs 0.528.
- On MAW intrinsic decomposition, SDM achieved albedo MSE 0.705 vs 0.924 for Li et al., with shading MSE 0.006 vs 0.021.
- In an AMT user study on object insertion, SDM was selected 0.322 of the time versus 0.348 for ground truth and 0.178 for Li et al.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Reduce training T from 1000 to 1 (ablation) | Diffusion model · OpenRooms FF test | N MSE: 2.236 -> 1.817 | Helped |
| Reduce training T from 1000 to 1 (ablation) | Diffusion model · OpenRooms FF test | N variance: 0.183 -> 0.0 | Hurt |
| Reduce training T from 1000 to 1 (ablation) | Diffusion model · OpenRooms FF test | D MSE: 0.065 -> 0.061 | Helped |
| Reduce training T from 1000 to 1 (ablation) | Diffusion model · OpenRooms FF test | A MSE: 0.418 -> 0.334 | Helped |
| Apply channel-wise noise scheduling in geometry→material→lighting order instead of uniform scheduling (ablation) | PDM · OpenRooms FF test | A_L MSE: 0.403 -> 0.376 | Helped |
| Apply channel-wise noise scheduling in reverse lighting→material→geometry order (ablation) | PDM · OpenRooms FF test | A_L MSE: 0.376 -> 0.414 | Hurt |
| Increase SDM iterative prediction from T=1 to T=4 (ablation) | SDM · OpenRooms FF test | A MSE: 0.318 -> 0.288 | Helped |
| Increase SDM iterative prediction from T=4 to T=7 (ablation) | SDM · OpenRooms FF test | A MSE: 0.288 -> 0.292 | Hurt |
Surprising findings
- Reducing the training timestep T from 1000 to 1 improves prediction accuracy in most cases at the cost of sample diversity, showing a trade-off between diversity and accuracy.
- RGB↔X achieves strong albedo prediction on MAW but reconstructs input images with the lowest accuracy.
Limitations
- Requires two separate models (PDM and SDM) rather than a unified model.
- Re-rendering error as an uncertainty metric may be imprecise.
- Does not exploit pretrained LDM priors, which may limit performance.
Experimental setup
Datasets: OpenRooms FF InteriorVerse MAW DIODE Garon et al. real-world dataset
Benchmarks: OpenRooms FF synthetic test set Garon et al. real-world images MAW DIODE
Metrics: MSE variance hue texture (LPIPS) WHDR AMT user preference
Baselines: Li et al. (OpenRooms) Zhu et al. (InteriorVerse) MAIR MAIR++ RGB↔X
Builds on: MAIR++ cosine noise scheduler v-prediction diffusion cascaded diffusion
Closest papers by meaning
- Diffusion Renderer: Neural Inverse and Forward Rendering with Video Diffusion Models (CVPR 2025 · similarity 0.90)
- Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion (CVPR 2025 · similarity 0.88)
- Intrinsic Image Diffusion for Indoor Single-view Material Estimation (CVPR 2024 · similarity 0.88)
- DiffuScene: Denoising Diffusion Models for Generative Indoor Scene Synthesis (CVPR 2024 · similarity 0.88)
- Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.88)
- Relay Diffusion: Unifying diffusion process across resolutions for image synthesis (ICLR 2024 · similarity 0.87)
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.