Research map / Image generation / Autoregressive and latent generation
DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection
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 DiffusionEdge, the first diffusion probabilistic model for general edge detection, running a decoupled DDM in autoencoder latent space with an adaptive FFT filter and a distilled uncertainty-aware cross-entropy loss to produce crisp edge maps without NMS post-processing.
Problem
Learning-based edge detectors struggle to produce edge maps that are both correct (accurate edge/non-edge labels) and crisp (thin, single-pixel-width edges) without heavy post-processing such as NMS, and diffusion models are costly to run in pixel space.
Approach
Apply a decoupled diffusion model in autoencoder latent space, add a learnable adaptive FFT filter to the denoising U-Net, and train with an uncertainty-aware weighted BCE loss passed to the latent code by skipping the autoencoder gradient (uncertainty distillation) so the diffusion output is crisp directly.
Key results
- On NYUDv2, compared to the second best (EDTER), ODS and OIS under CEval increase by 30.2% and 28.1% and AC by 65.1%, per the paper's abstract.
- On BSDS, DiffusionEdge reaches ODS 0.834 / OIS 0.848 under SEval with AC 0.476, versus UAED at 0.829 / 0.847 and AC 0.227.
- On BIPED, AC of 0.849 and CEval ODS 0.899 are reported, with edges almost all single-width without post-processing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Uncertainty Distillation without Adaptive FFT-filter (ablation) | DiffusionEdge · BSDS | ODS: 0.816 -> 0.831 | Helped |
| add Adaptive FFT-filter without Uncertainty Distillation (ablation) | DiffusionEdge · BSDS | ODS: 0.816 -> 0.825 | Helped |
| add Adaptive FFT-filter when Uncertainty Distillation is present (ablation) | DiffusionEdge with UD · BSDS | ODS: 0.831 -> 0.834 | Helped |
| add Adaptive FFT-filter without Uncertainty Distillation (ablation) | DiffusionEdge · BSDS | AC: 0.521 -> 0.461 | Hurt |
| use Swin instead of ResNet as image encoder (ablation) | DiffusionEdge · BSDS | ODS: 0.823 -> 0.834 | Helped |
| vary sampling steps from 1 to 50 (ablation) | DiffusionEdge with Swin · BSDS | ODS: 0.833; 0.835; 0.834; 0.834; 0.833; 0.833 | No clear effect |
Surprising findings
- On NYUDv2 the SEval ODS (0.761) is below EDTER (0.774), so the gains come mainly from no-post-processing crispness rather than SEval correctness.
- Using the autoencoder gradient (Baseline-A) performed worse than skipping it while using 1.5x more GPU memory.
Limitations
- Inference efficiency remains an open problem; sampling is slower than CNN detectors.
- Gains in SEval correctness on NYUDv2 are not consistently best (EDTER is higher on SEval ODS/OIS).
Experimental setup
Datasets: BSDS500 NYUDv2 Multicue BIPED
Benchmarks: BSDS500 NYUDv2 Multicue BIPED
Metrics: ODS OIS AC SEval CEval
Baselines: HED RCF BDCN PiDiNet EDTER UAED DexiNed AMH-Net
Builds on: decoupled diffusion model latent diffusion model RCF HED
Closest papers by meaning
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.87)
- DiP-GO: A Diffusion Pruner via Few-step Gradient Optimization (NeurIPS 2024 · similarity 0.87)
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.86)
- High-Precision Dichotomous Image Segmentation via Probing Diffusion Capacity (ICLR 2025 · similarity 0.86)
- EdgeDiff: Edge-aware Diffusion Network for Building Reconstruction from Point Clouds (CVPR 2025 · similarity 0.86)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · 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.