Research map / Image generation / Autoregressive and latent generation
Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis
Hila Chefer, Patrick Esser, Dominik Lorenz, Dustin Podell, Vikash Raja, Vinh Tong, Antonio Torralba, Robin Rombach
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 DINOv2-B with stronger DINO encoders (DINOv2-L, DINOv3-B, DINOv3-H+) in REPA | REPA · text-to-image FID vs training steps | FID | Hurt |
| use naive full masking or per-token diffusion forcing noise instead of uniform noise (ablation) | flow matching · noise scheduler comparison | FID | Hurt |
| add Dual-Timestep Scheduling alone without self-supervised loss (ablation) | vanilla flow matching · noise scheduler comparison | FID | Improved |
| add Self-Flow self-supervised objective with EMA teacher | SiT-XL/2 on ImageNet · ImageNet 256x256 (4M steps) | FID: 5.89 -> 5.70 (vs REPA) | Improved |
| add Self-Flow on top of RAE | RAE with SiT-XL · ImageNet 256x256 (1M steps) | FID: 3.24 -> 2.95 | Improved |
| add Self-Flow vs vanilla flow matching | text-to-image model · Text-to-image (1M steps) | FID: 4.08 -> 3.61 | Improved |
| add external alignment with video encoders (V-JEPA2, Depth Anything 3) | video flow model · Text-to-video (600K steps) | FVD: 50.95 -> 53.55 (V-JEPA2) | Hurt |
| add external alignment with MERT | audio flow model · Text-to-audio (350K steps) | FAD (CLAP): 148.874 -> 148.883 | No clear effect |
| scale model from 290M to 1B parameters | Self-Flow vs REPA · text-to-image scaling | CLIP score | Improved |
| remove representation loss Lrep (ablation) | Self-Flow · ImageNet class-to-image | FID: over 4 points worse without | Improved |
| add masking (dual-timestep) to representation alignment (ablation) | Self-Flow · ImageNet class-to-image | FID: over 1 point worse without | Improved |
| sample both timesteps from the full noise distribution instead of s in [t, t-0.2] (ablation) | Self-Flow · ImageNet class-to-image | FID | Improved |
Closest papers by meaning
- Flowing from Words to Pixels: A Noise-Free Framework for Cross-Modality Evolution (CVPR 2025 · similarity 0.89)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.88)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.88)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.88)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.88)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.88)
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.
