Research map / Image generation / Fast diffusion sampling
Going beyond Compositions, DDPMs Can Produce Zero-Shot Interpolations
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 DDPMs trained only on extreme (non-overlapping) subsets of CelebA-type data can, via multi-guidance classifier guidance, generate zero-shot interpolations such as mild smiles and intermediate ages/hair colors.
Problem
Whether diffusion models trained on strictly separated subsets of a latent factor's support (large gaps, e.g. only extreme smiles) can generate samples in the unseen intermediate region, i.e. zero-shot interpolation beyond mere composition of seen factors.
Approach
Train DDPMs with classifier guidance on extreme examples only, and use a multi-class guidance score over the extreme classes so that sampling is steered toward intermediate attribute values, producing mild attributes never seen in training.
Key results
- Models trained only on extreme CelebA smile examples generate mild smiles via multi-guidance; guidance strength 30 chosen as a good trade-off, with FID (full) 11.34 vs 8.14 unconditional.
- Interpolation degrades with 5k training samples while 30k and 10k perform similarly; memorization appears at 5k.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase classifier-guidance strength from 0.0 to 30 (ablation) | DDPM multi-guidance sampling · CelebA mild examples | Recall: 0.4755 -> 0.4353 | Hurt |
| Increase classifier-guidance strength from 0.0 to 30 (ablation) | DDPM multi-guidance sampling · CelebA mild examples | FID: 12.7215 -> 14.9275 | Hurt |
| Increase classifier-guidance strength from 0.0 to 10 (ablation) | DDPM multi-guidance sampling · CelebA mild examples | Precision: 0.7564 -> 0.8182 | Helped |
| Increase classifier-guidance strength to 100 (ablation) | DDPM multi-guidance sampling · CelebA full distribution | Recall: 0.484 -> 0.5383 | Mixed |
| Reduce training data to 30k or 10k examples rather than 60k (ablation) | DDPM multi-guidance · CelebA smile interpolation | KLD | No clear effect |
| Reduce training data to 5k examples (ablation) | DDPM multi-guidance · CelebA smile interpolation | KLD | Hurt |
| Replace the U-net with a diffusion transformer (ablation) | DDPM interpolation · CelebA smile interpolation | interpolation capability | No clear effect |
| Use classifier-free guidance instead of classifier guidance (ablation) | DDPM multi-guidance · CelebA smile interpolation | interpolation capability | No clear effect |
| Use the VLB objective (ablation) | DDPM · CelebA | sample realism | Hurt |
| Vary the noise schedule between cosine and linear, or use fixed versus learned variance (ablation) | DDPM interpolation · CelebA | interpolation capability | No clear effect |
| Use spherical linear interpolation instead of regular linear interpolation (ablation) | DDIM latent interpolation · CelebA smile interpolation | interpolation quality | Helped |
| Train the attribute classifier on synthetic green-lips examples (ablation) | DDPM sampling · CelebA green-lips extrapolation | green-lips generation | Helped |
Surprising findings
- Classifier guidance on green-lip synthetic data enables generation of green lips never seen in training (extrapolation).
- Interpolation capability persists with diffusion transformers, classifier-free guidance, and different noise schedules and variances.
Limitations
- Interpolation quality degrades with small training sets (5k), with memorization observed.
- FID values higher than state-of-the-art due to short training runs.
- Precision/recall metrics correlate weakly with manual assessment.
Experimental setup
Datasets: CelebA ImageNet
Benchmarks: CelebA
Metrics: FID precision recall KL divergence of attribute histograms MSE
Baselines: unconditional DDPM sampling classifier-free guidance
Builds on: DDPM Improved DDPM classifier guidance classifier-free guidance DDIM
Closest papers by meaning
- UDPM: Upsampling Diffusion Probabilistic Models (NeurIPS 2024 · similarity 0.89)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.88)
- On Inference Stability for Diffusion Models (AAAI 2024 · similarity 0.88)
- An Edit Friendly DDPM Noise Space: Inversion and Manipulations (CVPR 2024 · similarity 0.88)
- Diffusion Bridge Implicit Models (ICLR 2025 · 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.