Research map / Image generation / Editing and style transfer
Contrastive Diffusion Alignment: Learning Structured Latents for Controllable Generation
Ruchi Sandilya, Sumaira Perez, Charles Lynch, Lindsay Victoria, Benjamin Zebley, Derrick Matthew Buchanan, Mahendra T. Bhati, Nolan Williams, Timothy Spellman, FAITH M. GUNNING, Conor Liston, Logan Grosenick
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 |
|---|---|---|---|
| interpolate in the contrastive embedding C instead of raw diffusion latent Z (TEX-1) (ablation) | TEX-1 · Fluid interpolation | PSNR: 29.28 -> 32.98 | Improved |
| use C-space instead of Z-space for LSTM RMSE (ablation) | LSTM · Fluid interpolation | RMSE: 4.04 -> 1.07 | Improved |
| use C-space instead of Z-space for TEX-2 (ablation) | TEX-2 · Fluid interpolation | PSNR: 35.29 -> 35.70 | Improved |
| linear and spherical interpolation (Lerp/Slerp) in either space (ablation) | Lerp/Slerp · Fluid interpolation | PSNR: 28.16 -> 28.27 | No clear effect |
| second-order Taylor term (TEX-2) instead of first-order (TEX-1) (ablation) | TEX extrapolation · Ca2+ interpolation (C-space) | PSNR: 36.13 -> 38.58 | Improved |
| n-step-ahead prediction with spline vs LSTM vs TEX-2 in C-space (ablation) | ConDA · Ca2+ n-step-ahead | PSNR: 35.11 (LSTM) / 34.92 (TEX-2) / 35.00 (Spline) | No clear effect |
| train SVM on ConDA embedding rather than raw diffusion latent (E-field responder) (ablation) | SVM-RBF · Responder vs non-responder | F1: 0.30 -> 0.63 | Improved |
| train SVM on ConDA embedding rather than raw diffusion latent (flow) (ablation) | SVM-RBF · Steady vs unsteady flow | accuracy: 0.68 -> 0.87 | Improved |
| use contrastive ConDA embedding instead of PCA (ablation) | ConDA · Fluid held-out segment prediction | PSNR: 31.82 -> 33.90 | Improved |
| use contrastive ConDA embedding instead of beta-VAE (ablation) | ConDA · Fluid held-out segment prediction | RMSE: 7.47 -> 7.60 | Mixed |
| use contrastive ConDA embedding instead of beta-VAE (ablation) | ConDA · DISFA held-out segment prediction | PSNR: 35.76 -> 36.26 | Improved |
| kNN decoder lifting instead of tuned MLP decoder (ablation) | ConDA · multiple datasets | reconstruction | Improved |
Closest papers by meaning
- NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 2024 · similarity 0.86)
- AHPA: Adaptive Hierarchical Prior Alignment for Diffusion Transformers (NeurIPS 2026 · similarity 0.85)
- Diff-CA: Separating Common and Salient Factors with Diffusion Models (NeurIPS 2026 · similarity 0.85)
- Contrast encodes inductive bias: separating slow noise from dynamics in predictive representation learning (NeurIPS 2026 · similarity 0.85)
- Steering Optimisation Trajectories in Diffusion Representation Learning (NeurIPS 2026 · similarity 0.85)
- FreeControl: Efficient, Training-Free Structural Control via One-Step Attention Extraction (NeurIPS 2025 · similarity 0.85)
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.
