Research map / Image generation / Autoregressive and latent generation
Learning on the Manifold: Unlocking Standard Diffusion Transformers with Representation Encoders
Amandeep Kumar, Vishal Patel
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 |
|---|---|---|---|
| replace Euclidean flow matching with Riemannian flow matching (SLERP geodesic) (ablation) | DiT-B/1 on DINOv2-B · ImageNet 256 | FID: 24.32 -> 7.06 | Improved |
| project noise and latents to the unit sphere only (ablation) | DiT-B/1 on DINOv2-B · ImageNet 256 | FID: 24.32 -> 21.99 | No clear effect |
| add Jacobi regularization on top of RFM (ablation) | DiT-B/1 on DINOv2-B · ImageNet 256 | FID: 7.06 -> 6.77 | Improved |
| train RJF longer (80 to 200 epochs) (ablation) | DiT-B/1 RJF · ImageNet 256 | FID: 6.77 -> 4.95 | Improved |
| add RJF to LightningDiT-XL (DINOv2-B) | LightningDiT-XL · ImageNet 256 | FID: 4.28 -> 3.62 | Improved |
| add RJF to DDT-XL/1 | DDT-XL · ImageNet 256 | FID: 6.55 -> 5.82 | Improved |
| add RJF to width-scaled DiT^DH | DiT^DH · ImageNet 256 | FID: 6.33 -> 6.20 | Improved |
| add RJF with SigLIP latents | DiT-B/1 · ImageNet 256 (SigLIP) | FID: 130.21 -> 10.39 | Improved |
| add RJF with MAE latents | DiT-B/1 · ImageNet 256 (MAE) | FID: 50.48 -> 19.82 | Improved |
| increase inference projection radius from original norm to R=45 (ablation) | DiT-B/1 RJF · ImageNet 256 | FID: 7.79 -> 6.77 | Improved |
| compare RJF with guidance to SFD at 80 epochs | DiT-XL RJF · ImageNet 256 (guidance) | FID: 2.81 vs 1.30 (SFD) | Hurt |
| mask the radial loss and train only angular component (width sweep) (ablation) | DiT-S single image overfit · DINOv2 tokens d=768 | angular loss convergence | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- REPA-E — REPA-E: Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers (ICCV 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Flow Matching on General Geometries (ICLR 2024 · similarity 0.90)
- Metric Flow Matching for Smooth Interpolations on the Data Manifold (NeurIPS 2024 · similarity 0.88)
- Generative Modeling on Manifolds Through Mixture of Riemannian Diffusion Processes (ICML 2024 · similarity 0.88)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.87)
- Flow-Corrected Shape Optimization: Taming Manifold Drift in High-Dimensional 3D Models (NeurIPS 2026 · similarity 0.87)
- Riemannian Consistency Model (NeurIPS 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.
