Research map / Image generation / Fast diffusion sampling
Subspace Diffusion Generative Models
Bowen Jing, Gabriele Corso, Renato Berlinghieri, Tommi Jaakkola
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 |
|---|---|---|---|
| add 16x16 subspace diffusion | NCSN++ · CIFAR-10 | FID: 2.38 -> 2.29 | Improved |
| add 16x16 subspace diffusion | NCSN++ (deep) · CIFAR-10 | FID: 2.20 -> 2.17 | Improved |
| add 16x16 subspace diffusion | DDPM++ · CIFAR-10 | FID: 2.61 -> 2.60 | No clear effect |
| add 16x16 subspace diffusion | DDPM++ (deep) · CIFAR-10 | FID: 2.41 -> 2.40 | No clear effect |
| add 16x16 subspace diffusion (runtime) | NCSN++ · CIFAR-10 | relative runtime: 100% -> 66% | Improved |
| raise divergence threshold from 3e-3 to 3e-2 (more diffusion in subspace) (ablation) | NCSN++ 16x16 · CIFAR-10 | FID: 2.29 -> 2.67 | Hurt |
| raise divergence threshold from 1e-4 to 3e-3 (ablation) | NCSN++ 16x16 · CIFAR-10 | FID: 2.45 -> 2.29 | Improved |
| raise divergence threshold from 1e-4 to 3e-2 (ablation) | NCSN++ 8x8 · CIFAR-10 | FID: 2.41 -> 3.42 | Hurt |
| raise divergence threshold from 1e-4 to 3e-2 (ablation) | DDPM++ 16x16 · CIFAR-10 | FID: 2.61 -> 5.68 | Hurt |
| use 8x8 instead of 16x16 subspace at threshold 1e-3 (ODE) (ablation) | DDPM++ (deep) · CIFAR-10 ODE | FID: 3.17 -> 3.53 | Hurt |
| add subspace diffusion with ODE sampling (16x16, 1e-3) | DDPM++ (deep) · CIFAR-10 ODE | FID: 2.95 -> 3.17 | Hurt |
| add subspace diffusion with ODE sampling (16x16, 1e-3) | DDPM++ (deep) · CIFAR-10 ODE | NLL: 2.995 -> 2.997 | No clear effect |
Closest papers by meaning
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Evaluating the design space of diffusion-based generative models (NeurIPS 2024 · similarity 0.90)
- Dimension-free Score Matching and Time Bootstrapping for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models (ICLR 2024 · similarity 0.89)
- Training Diffusion-based Generative Models with Limited Data (ICML 2025 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · 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.
