Research map / Image generation / Autoregressive and latent generation
D2C: Diffusion-Decoding Models for Few-Shot Conditional Generation
Abhishek Sinha, Jiaming Song, Chenlin Meng, Stefano Ermon
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| add contrastive SSL (MoCo-v2) loss to the VAE/diffusion latent model | D2C vs NVAE/DDIM · CIFAR-10 | linear accuracy: 22.5 -> 76.02 | Improved |
| add diffusion prior over VAE latents (D2C vs NVAE) | D2C · CelebA-64 | FID: 13.48 -> 5.7 | Improved |
| use D2C instead of DDIM for unconditional generation | D2C · CIFAR-10 | FID: 4.16 -> 10.15 | Hurt |
| use D2C instead of DDIM | D2C · CelebA-HQ-256 | FID: 25.6 -> 18.74 | Improved |
| use D2C latent classifier with rejection sampling instead of DDIM | D2C · CelebA-64 few-shot binary Male | FID: 38.38 -> 13.44 | Improved |
| use D2C instead of NVAE for few-shot conditional generation | D2C · CelebA-64 PU Male | FID: 42.78 -> 16.39 | Improved |
| use D2C instead of naive baseline | D2C · CelebA-64 binary Non-Blond | FID: 3.77 -> 8.94 | Hurt |
| fewer diffusion steps (100 -> 10) (ablation) | D2C · CIFAR-10 | FID: 10.15 -> 17.71 | Hurt |
| D2C vs StyleGAN2 for image manipulation | D2C · CelebA-256 AMT | preference %: 51.5% blond, 60.8% red lipstick | Improved |
| replace StyleGAN2 latent optimization with D2C encoder | D2C · manipulation latent time | seconds: 8 seconds -> 0.013 seconds | Improved |
| vary contrastive loss weight lambda (ablation) | D2C · unconditional | FID | No clear effect |
| train contrastive SSL alone without L_D2 (ablation) | contrastive SSL · CIFAR-10 | linear accuracy: 78.3% vs 76.02 | Mixed |
Closest papers by meaning
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.89)
- A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- Referee Can Play: An Alternative Approach to Conditional Generation via Model Inversion (ICML 2024 · similarity 0.87)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Nested Diffusion Models Using Hierarchical Latent Priors (CVPR 2025 · similarity 0.87)
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.
