Research map / Image generation / Fast diffusion sampling
EVODiff: Entropy-aware Variance Optimized Diffusion Inference
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 EVODiff, an information-theoretic, entropy-aware variance-optimized second-order solver for data-prediction diffusion inference that reduces FID versus DPM-Solver++ and UniPC without reference trajectories.
Problem
Slow diffusion model inference lacks an information-theoretic foundation; existing gradient-based ODE solvers ignore information transmission efficiency and conditional entropy in reverse transitions.
Approach
Frames denoising as conditional entropy reduction in reverse transitions, shows conditional entropy is tied to conditional variance, and optimizes the step-combination coefficients (zeta, eta) in closed form from evolution state differences, giving a reference-free second-order solver that favors data prediction.
Key results
- On CIFAR-10 (EDM) at 10 NFE, EVODiff reduces FID from 5.10 (DPM-Solver++) to 2.78, a 45.5% reduction.
- On ImageNet-256 (ADM), EVODiff reaches FID 7.48 at 15 NFE vs 7.80 for DPM-Solver++ at 15 NFE, matching DPM-Solver++ quality at 20 NFE with fewer steps (the abstract's 25% NFE saving).
- On LSUN-Bedrooms latent diffusion at 5 NFE, EVODiff FID 7.912 vs UniPC 13.969 (43.4% gain).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace DPM-Solver++ with EVODiff | EDM · CIFAR-10, 10 NFE | FID: 5.10 -> 2.78 | Helped |
| Replace DPM-Solver++ with EVODiff | EDM · FFHQ-64, 10 NFE | FID: 6.81 -> 5.31 | Helped |
| Replace DPM-Solver-v3 with EVODiff | ADM · ImageNet-256, 20 NFE | FID: 7.42 -> 7.25 | Helped |
| Replace UniPC(3m) with EVODiff | Latent diffusion model · LSUN-Bedrooms, 5 NFE | FID: 13.969 -> 7.912 | Helped |
| Replace DPM++ with EVODiff | Latent diffusion model · LSUN-Bedrooms, 5 NFE | Generation Time (s): 3577.6 -> 3488.4 | Helped |
| Replace DPM-Solver++ with EVODiff | Stable Diffusion v1.4 · Text-to-image, NFN=10 | CLIP Average: 33.07 -> 33.79 | Helped |
| Replace DPM-Solver++ with EVODiff | Stable Diffusion v1.5 · Text-to-image, NFN=25 | Aesthetic Average: 5.79 -> 5.81 | Helped |
| Add the RE-based formulation to the baseline second-order solver (ablation) | CIFAR-10 diffusion inference · CIFAR-10, 10 NFE | FID: 5.10 -> 2.98 | Helped |
| Add final evolution-state-driven parameter optimization to the RE-based formulation (ablation) | CIFAR-10 diffusion inference · CIFAR-10, 10 NFE | FID: 2.98 -> 2.78 | Helped |
| Use balanced ζ_i (ablation) | ADM diffusion inference · ImageNet-256, 10 NFE | FID: 8.68 -> 8.57 | Helped |
| Refine ζ_i instead of using balanced ζ_i (ablation) | ADM diffusion inference · ImageNet-256, 10 NFE | FID: 8.57 -> 8.23 | Helped |
| Use EVODiff with refined step-size ratio instead of refined ζ_i alone (ablation) | ADM diffusion inference · ImageNet-256, 10 NFE | FID: 8.23 -> 8.14 | Helped |
Surprising findings
- Reported FID gains on LSUN-Bedrooms come with only about 2.5% to 5% generation-time savings, and some table entries show EVODiff slightly worse than baselines (e.g., 3.756 vs 3.963 at 8 NFE is better, but FFHQ IS at 5 NFE is lower).
Limitations
- Currently relies on data-prediction parameterization.
- Theoretical analysis assumes independence of model noise estimates across timesteps and Gaussian reverse transitions under isotropic assumptions.
- Gains over strong baselines are modest at higher NFE (e.g., CIFAR-10 at 20 NFE, 2.06 vs 2.33 for baseline).
Experimental setup
Datasets: CIFAR-10 CelebA-64 FFHQ-64 ImageNet-64 ImageNet-256 LSUN-Bedrooms
Benchmarks: CIFAR-10 FFHQ-64 ImageNet-256 LSUN-Bedrooms ImageNet-64 CelebA-64
Metrics: FID IS CLIP score Aesthetic score generation time
Baselines: Heun DPM-Solver++ UniPC DPM-Solver-v3 DEIS DDIM
Builds on: DPM-Solver DPM-Solver++ DDIM EDM UniPC
Closest papers by meaning
- On Inference Stability for Diffusion Models (AAAI 2024 · similarity 0.90)
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.90)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.90)
- Reverse Transition Kernel: A Flexible Framework to Accelerate Diffusion Inference (NeurIPS 2024 · similarity 0.90)
- Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.90)
- D^2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models (AAAI 2025 · 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.