Research map / Probabilistic and scientific modeling / Variational and generative sampling
Latent Stochastic Interpolants
Saurabh Singh, Dmitry Lagun
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 |
|---|---|---|---|
| train latent SI jointly with encoder and decoder (LSI) instead of observation-space SI | Stochastic interpolants · ImageNet 256x256 | FID: 3.87 -> 3.91 | No clear effect |
| train in latent space rather than pixel space | Stochastic interpolants · ImageNet 256x256 | GFLOPs per forward pass of generative model: 1288 -> 450 | Improved |
| joint training (beta > 0) instead of independent encoder-decoder (beta -> 0) (ablation) | LSI · ImageNet 128x128 | FID: 4.53 -> 3.75 | Improved |
| increase beta beyond the best value (ablation) | LSI · ImageNet 128x128 | FID | Hurt |
| joint training when shifting capacity from latent model to encoder/decoder (k=6) (ablation) | LSI · ImageNet 128x128 | FID: 4.87 -> 3.96 | Improved |
| use constant or learned encoder noise instead of a deterministic encoder (ablation) | LSI · ImageNet 128x128 | FID | Improved |
| use the proposed parameterization instead of denoising, original flow or noise prediction (ablation) | LSI · ImageNet 128x128 FID @1K epochs | FID: 4.28 -> 3.76 | Improved |
| use Gaussian prior instead of Uniform, Laplacian or Gaussian Mixture p0 (ablation) | LSI · ImageNet 128x128 FID @1K epochs | FID: 4.26 (GMM) -> 3.76 | Improved |
| move first/last k=9 blocks from latent model into encoder/decoder (ablation) | LSI · ImageNet 128x128 | FID: 3.76 -> 4.61 | Hurt |
| use fixed constant encoder noise instead of learned encoder variance (ablation) | LSI · ImageNet 128x128 | FID | Improved |
Closest papers by meaning
- Stochastic Interpolants via Conditional Dependent Coupling (NeurIPS 2026 · similarity 0.90)
- Multitask Learning with Stochastic Interpolants (NeurIPS 2025 · similarity 0.88)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.87)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.87)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · 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.
