Research map / Image generation / Fast diffusion sampling
Deep Equilibrium Approaches to Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace sequential DDIM sampling with joint DEQ-DDIM sampling | DDIM · CIFAR10, T=1000 | time: 20.16s -> 2.91s | Helped |
| Replace sequential DDIM sampling with joint DEQ-DDIM sampling | DDIM · CelebA, T=500 | FID: 3.66 -> 2.92 | Helped |
| Replace sequential DDIM sampling with joint DEQ-DDIM sampling | DDIM · LSUN Bedroom, T=25 | FID: 8.76 -> 8.73 | No clear effect |
| Replace sequential DDIM sampling with joint DEQ-DDIM sampling | DDIM · LSUN Church, T=25 | time: 1.68s -> 3.99s | Hurt |
| Use DEQ-DDIM instead of sequential DDIM for latent inversion | DDIM inversion · CIFAR10, T=10 | minimum inversion loss: 2.59 ± 3.67 -> 0.68 ± 0.32 | Helped |
| Use DEQ-DDIM instead of sequential DDIM for latent inversion | DDIM inversion · CelebA, T=20 | minimum inversion loss: 14.13 ± 5.04 -> 1.03 ± 0.37 | Helped |
| Use DEQ-DDIM instead of sequential DDIM for latent inversion | DDIM inversion · LSUN Bedroom, T=10 | minimum inversion loss: 1114.49 ± 795.86 -> 36.37 ± 22.86 | Helped |
| Use DEQ-DDIM instead of sequential DDIM for latent inversion | DDIM inversion · LSUN Church, T=10 | minimum inversion loss: 1674.68 ± 1432.54 -> 47.94 ± 24.78 | Helped |
| Use inexact one-step gradients instead of exact gradients (ablation) | DEQ-DDIM inversion · CelebA, chain length 20 | convergence rate: + faster convergence at learning rate 0.01; no numeric value reported | Helped |
| Initialize all latent states with x_T instead of zeros (ablation) | DEQ-DDIM · CIFAR10, chain length 50 | convergence rate: up to 3× | Helped |
| Increase the chain length during inversion (ablation) | DEQ-DDIM · CIFAR10 and CelebA | convergence rate | No clear effect |
| Increase sampling steps after inversion (ablation) | Baseline DDIM inversion · LSUN Bedroom | reconstruction quality | No clear effect |
Closest papers by meaning
- Diffusion Bridge Implicit Models (ICLR 2025 · similarity 0.93)
- Deep Equilibrium Diffusion Restoration with Parallel Sampling (CVPR 2024 · similarity 0.92)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.91)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.91)
- Fixed Point Diffusion Models (CVPR 2024 · similarity 0.91)
- Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models (NeurIPS 2026 · similarity 0.90)
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.