Research map / Image generation / Fast diffusion sampling
Increasing the Utility of Synthetic Images through Chamfer Guidance
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 Chamfer Guidance, a training-free inference-time guidance that uses a few real exemplar images and a DINOv2-space Chamfer distance to boost diversity and quality of class-conditional/text-to-image latent diffusion samples used as synthetic training data.
Problem
Conditional image generators trade diversity for quality (high CFG), limiting their utility as synthetic training data, and existing guidance methods ignore the distribution shift between synthetic and real data.
Approach
Use the symmetric Chamfer distance between DINOv2 features of a batch of generated samples (via DDIM-approximated denoised estimates) and a handful of real exemplar images as a differentiable guidance signal, applied every few denoising steps without fine-tuning and without an unconditional model.
Key results
- With SD 3.5 and 2 exemplars, Chamfer Guidance reaches 96.4% precision and 86.4% coverage on ImageNet-1k (97.5% and 92.7% with 32 exemplars), per the abstract.
- On GeoDE, DINOv2-based Chamfer Guidance with 4 exemplars improves average F1 by about 7% over c-VSG in Inception space.
- Classifiers trained on synthetic data from Chamfer Guidance gain up to 15% in-distribution accuracy over CFG sampling, and up to 16% on ImageNet-Sketch out-of-distribution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Chamfer Guidance with 32 real exemplars instead of Chamfer fine-tuning | LDM-3.5 · ImageNet-1k | F1 (P, C): 0.731 -> 0.950 | Helped |
| Use Chamfer Guidance with 32 real exemplars instead of CFG sampling | LDM-1.5 synthetic-only classifier · ImageNet-1k validation | accuracy: 47.67 -> 54.91 | Helped |
| Use Chamfer Guidance with 32 real exemplars instead of CFG sampling | LDM-3.5 synthetic-only classifier · ImageNet-Sketch | accuracy: 17.60 -> 34.44 | Helped |
| Use DINOv2 Chamfer Guidance with 4 exemplars instead of c-VSG with 4 exemplars | LDM-1.5 · GeoDE, DINOv2 metrics | F1 (P, C): 0.184 -> 0.304 | Helped |
| Use DINOv2 Chamfer Guidance with 4 exemplars instead of c-VSG with 4 exemplars | LDM-1.5 · DollarStreet | F1 (P, C): 0.517 -> 0.552 | Helped |
| Increase the number of real exemplars from 2 to 32 | Chamfer Guidance on ImageNet-1k · ImageNet-1k, LDM-3.5 | Coverage: 0.864 -> 0.927 | Helped |
| Set CFG scale to ω=2.0 rather than ω=1.0 (ablation) | Chamfer Guidance on LDM-1.5 with k=32 · ImageNet-1k | F1 (P, C): 0.899 -> 0.931 | Helped |
| Set CFG scale to ω=7.5 rather than ω=2.0 (ablation) | Chamfer Guidance on LDM-1.5 with k=32 · ImageNet-1k | FID: 8.935 -> 14.388 | Hurt |
| Set Chamfer guidance strength to γ=0.07 rather than γ=0.05 (ablation) | Chamfer Guidance on LDM-1.5 with k=32 · ImageNet-1k | F1 (P, C): 0.914 -> 0.931 | Helped |
| Set Chamfer guidance strength to γ=0.07 rather than γ=0.05 (ablation) | Chamfer Guidance on LDM-1.5 with k=32 · ImageNet-1k | FID: 8.906 -> 8.935 | Hurt |
| Use Chamfer Guidance rather than CFG-based sampling | LDM-3.5 · ImageNet-1k sampling | FLOPs: 31% | Helped |
| Add real training data to Chamfer-generated synthetic data | LDM-3.5 classifier · ImageNet-R and ImageNet-A | accuracy | Hurt |
Surprising findings
- Reference-free guidance methods and training-based fine-tuning (vanilla and Chamfer fine-tuning) do not improve when more real exemplars are added, while Chamfer Guidance does.
- Newer, higher-quality SD 3.5 generations yield lower downstream classifier utility than SD 1.5 without guidance, which Chamfer Guidance partly reverses.
Limitations
- Designed for class-conditional models; does not directly support text-to-image generation.
- Evaluation relies on automated feature-based metrics that inherit dataset and feature-extractor biases.
- Requires real exemplar images at inference time.
Experimental setup
Datasets: ImageNet-1k GeoDE DollarStreet ImageNet-V2 ImageNet-Sketch ImageNet-R ImageNet-A
Benchmarks: ImageNet-1k GeoDE DollarStreet ImageNet-V2 ImageNet-Sketch ImageNet-R ImageNet-A
Metrics: F1 (precision, coverage) precision coverage density recall FDD FID CLIPScore top-1 accuracy
Baselines: CFG APG CADS Limited Interval guidance Particle Guidance c-VSG vanilla fine-tuning ReFL-based Chamfer fine-tuning Semantic Guidance FG CLIP Textual Inversion Paraphrasing
Builds on: classifier-free guidance c-VSG ReFL DDIM Vendi Score
Closest papers by meaning
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.88)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.87)
- Improved Training Technique for Shortcut Models (NeurIPS 2025 · similarity 0.87)
- Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.86)
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.86)
- REG: Rectified Gradient Guidance for Conditional Diffusion Models (ICML 2025 · similarity 0.86)
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.