Research map / Image generation / Fast diffusion sampling
Inverting Data Transformations via Diffusion Sampling
Jinwoo Kim, Sékou-Oumar Kaba, Jiyun Park, Seunghoon Hong, Siamak Ravanbakhsh
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 TIED test-time inversion with classifier-logit energy | ResNet18 on MNIST · affNIST test, Aff(2,R) | accuracy (%): 55.48 -> 82.64 | Improved |
| use VAE ELBO energy instead of classifier-logit energy (ablation) | TIED · affNIST test, Aff(2,R) | accuracy (%): 82.64 -> 96.95 | Improved |
| add diffusion at high noise levels (TIED) instead of gradient-only Langevin or LieLAC sampling (ablation) | Trivialized energy-based sampling · affNIST test, Aff(2,R), VAE energy | accuracy (%): LieLAC 94.38 -> TIED 96.95; Langevin 74.72 -> 96.95 | Improved |
| use TIED sampling vs trivialized kinetic Langevin on SO(10) | Energy-based Lie group sampler · SO(10) synthetic energy -10 X_{1,1}^2 | sampling steps to convergence: 100 diffusion steps vs 100,000 Langevin steps | Improved |
| add TIED to pretrained DeepONet on transformed 1D heat equation | DeepONet · 1D heat equation, SL(2,R) ⋉ H(1,R) | relative L2 error: 0.690 -> 0.043 | Improved |
| add TIED to DeepONet on 1D Burgers equation | DeepONet · 1D Burgers equation, SL(2,R) ⋉ (R2,+) | relative L2 error: 0.867 -> 0.178 | Improved |
| add TIED to DeepONet trained with data augmentation | DeepONet + data aug. · 1D heat equation, transformed | relative L2 error: 0.081 -> 0.052 | Improved |
| use FoCal Bayesian optimization on heat equation | DeepONet + FoCal · 1D heat equation, SL(2,R) ⋉ H(1,R) | relative L2 error: 0.690 -> 1.948 | Hurt |
| TIED with classifier energy versus specialized equivariant networks | ResNet18 + TIED · affNIST/homNIST | accuracy (%): VAE energy: 96.95 vs affConv 95.08; 97.43 vs homConv 95.71; classifier energy 82.64 and 89.84 | Mixed |
| use gradient-based energy methods with classifier energy under large affine transformations | Kinetic Langevin, LieLAC · affNIST, classifier energy | accuracy (%): Langevin 53.98, LieLAC 73.79 vs ResNet18 55.48 | Hurt |
Closest papers by meaning
- Lie Group Decompositions for Equivariant Neural Networks (ICLR 2024 · similarity 0.89)
- Diffusion Generative Modeling on Lie Group Representations (NeurIPS 2025 · similarity 0.87)
- DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.87)
- Learning Infinitesimal Generators of Continuous Symmetries from Data (NeurIPS 2024 · similarity 0.87)
- Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.86)
- Improved sampling via learned diffusions (ICLR 2024 · 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.
