Research map / Image generation / Fast diffusion sampling
Elucidating the design space of classifier-guided diffusion generation
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.
Shows that off-the-shelf ImageNet classifiers (ResNet) can guide DDPM, EDM and DiT training-free to beat fine-tuned classifier guidance and classifier-free guidance, using calibration-driven pre-conditioning (denoised-sample input, Softplus smoothing, joint/marginal logit temperatures, sine guidance schedule).
Problem
Classifier guidance and classifier-free guidance need costly training with labeled data and cannot easily adapt to new conditions, while training-free guidance with pretrained discriminative models has lagged in quantitative performance.
Approach
Use calibration (integral ECE along the reverse process) as a guideline to pre-condition off-the-shelf classifiers: feed predicted clean samples, smooth activations with Softplus, strengthen joint vs marginal logit guidance via separate temperatures, and add a sine-shaped guidance schedule.
Key results
- Off-the-shelf ResNet-101 guidance on DDPM ImageNet 128x128 reaches FID 2.19, beating fine-tuned classifier guidance (2.97) and classifier-free diffusion (2.43).
- On EDM ImageNet 64x64 (36 steps), ResNet-101 guidance lowers FID from 2.35 to 2.22; with DiT ImageNet 256x256, FID improves from 2.27 to 2.12.
- Calibration-related ablations show FID correlates with integral ECE (fine-tuned 0.51/8.27, ResNet 0.35/7.17, combined 0.28/6.94).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use calibrated off-the-shelf ResNet guidance instead of the fine-tuned classifier | DDPM · ImageNet 128x128 | FID: 2.97 -> 2.19 | Helped |
| Add off-the-shelf ResNet guidance to the diffusion baseline | DDPM · ImageNet 128x128 | FID: 5.91 -> 2.19 | Helped |
| Add off-the-shelf ResNet101 classifier guidance | EDM · ImageNet 64x64, 36 ODE steps | FID: 2.35 -> 2.22 | Helped |
| Add off-the-shelf ResNet101 classifier guidance | EDM · ImageNet 64x64, 18 ODE steps | FID: 2.54 -> 2.35 | Helped |
| Add off-the-shelf ResNet101 classifier guidance | EDM · ImageNet 64x64, 10 ODE steps | FID: 3.64 -> 3.38 | Helped |
| Add off-the-shelf ResNet101 guidance to classifier-free sampling | DiT · ImageNet 256x256 | FID: 2.27 -> 2.12 | Helped |
| Add off-the-shelf ResNet101 guidance to classifier-free sampling | DiT · ImageNet 256x256 | Precision: 0.828 -> 0.817 | Hurt |
| Add off-the-shelf ResNet101 guidance to classifier-free sampling | DiT · ImageNet 256x256 | Recall: 0.57 -> 0.59 | Helped |
| Increase recurrent guidance from 1 to 2 steps (ablation) | ResNet-guided DDPM · ImageNet 128x128 | FID: 7.17 -> 7.06 | No clear effect |
| Use predicted denoised samples instead of reverse-process samples as classifier input (ablation) | ResNet-guided DDPM · ImageNet 128x128 | FID: 8.61 -> 7.17 | Helped |
| Replace ReLU with smoother SoftPlus activation (ablation) | ResNet-guided DDPM · ImageNet 128x128 | FID: 7.17 -> 6.61 | Helped |
| Reduce marginal logit temperature τ₂ from 1.0 to 0.5 while fixing τ₁=1 (ablation) | ResNet-guided DDPM · ImageNet 128x128 | FID: 6.20 -> 5.27 | Helped |
Surprising findings
- Off-the-shelf ResNet is more calibrated than the fine-tuned time-dependent classifier at high noise levels.
- Increasing recurrent guidance steps of an off-the-shelf classifier does not meaningfully improve FID (7.17, 7.06, 7.14) while adding sampling time.
Limitations
- Evaluation is restricted to ImageNet class-conditional generation with ResNet classifiers; text-to-image is only demonstrated qualitatively with CLIP.
- Hyperparameters (Softplus beta, temperatures, sine factor) are tuned per model and sampler setting.
- Calibration proposition relies on smoothness assumptions that are only loosely connected to the practical design.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 128x128 ImageNet 64x64 ImageNet 256x256
Metrics: FID Precision Recall ECE CLIP score
Baselines: classifier guidance (fine-tuned) classifier-free guidance universal guidance recurrent guidance GLIDE EDM baseline DiT baseline
Builds on: classifier guidance classifier-free guidance universal guidance EDM DiT
Closest papers by meaning
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.92)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.92)
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.91)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.91)
- Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.91)
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.91)
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.