Research map / Probabilistic and scientific modeling / Variational and generative sampling
Flowing Backwards: Improving Normalizing Flows via Reverse Representation Alignment
Yang Chen, Xiaowei Xu, Shuai Wang, Chenhui Zhu, Ruxue Wen, Xubin Li, Tiezheng Ge, Limin Wang
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 reverse-pass representation alignment (R-REPA) to TARFlow | TARFlow · ImageNet 64x64 (1M iters) | FID-4096: 11.76 -> 11.25 | Improved |
| add R-REPA to TARFlow | TARFlow · ImageNet 64x64 (1M iters) | Accuracy (%): 39.97 -> 57.02 | Improved |
| add R-REPA to Latent-TARFlow | Latent-TARFlow · ImageNet 256x256 latent (1M iters) | Accuracy (%): 40.22 -> 56.24 | Improved |
| add R-REPA to Latent-TARFlow | Latent-TARFlow · ImageNet 256x256 (50K samples) | FID: 5.15 -> 4.95 | Improved |
| add R-REPA to Latent-TARFlow | Latent-TARFlow · ImageNet 256x256 (50K samples) | sFID: 6.78 -> 6.89 | Hurt |
| use patch size 1 in addition to R-REPA (ablation) | Latent-TARFlow + R-REPA · ImageNet 256x256 (50K samples) | FID: 4.95 -> 4.18 | Improved |
| use forward-graph alignment instead of none (ablation) | TARFlow · ImageNet 64x64 (400k iters, ablation) | sFID: 33.79 -> 37.97 | Hurt |
| use reverse instead of detach strategy on blocks 7&8 (ablation) | TARFlow alignment · ImageNet 64x64 (400k iters, ablation) | FID-4096: 12.12 -> 11.93 | Improved |
| align blocks 7&8 rather than blocks 1&2 (Detach) (ablation) | TARFlow alignment · ImageNet 64x64 (400k iters, ablation) | FID-4096: 12.73 -> 12.12 | Improved |
| align blocks 1&2 rather than 7&8 (Detach) (ablation) | TARFlow alignment · ImageNet 64x64 (400k iters, ablation) | Accuracy (%): 55.14 -> 61.63 | Improved |
| move alignment to deeper layer (2nd to 6th) (ablation) | R-REPA blocks 7&8 · ImageNet 64x64 (400k iters, ablation) | FID-4096: 11.93 -> 11.71 | Improved |
| add R-REPA to TARFlow | TARFlow · ImageNet 64x64 (50K samples) | FID: 4.21 -> 3.69 | Improved |
Closest papers by meaning
- Bidirectional Normalizing Flow: From Data to Noise and Back (CVPR 2026 · similarity 0.91)
- SimFlow: Simplified and End-to-End Training of Latent Normalizing Flows (ECCV 2026 · similarity 0.89)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.88)
- An Error Analysis of Flow Matching for Deep Generative Modeling (ICML 2025 · similarity 0.87)
- Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models (NeurIPS 2026 · similarity 0.86)
- JanusFlow: Harmonizing Autoregression and Rectified Flow for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.86)
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.
