Research map / Image generation / Autoregressive and latent generation
MeanFlow Transformers with Representation Autoencoders
Zheyuan Hu, Chieh-Hsin Lai, Ge Wu, Yuki Mitsufuji, Stefano Ermon
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 |
|---|---|---|---|
| run MeanFlow in RAE latent space with DiTDH instead of SD-VAE with DiT/SiT (ablation) | MeanFlow, MFD, unguided · ImageNet 256 1-step | FID: 5.94 -> 2.03 | Improved |
| train with MeanFlow distillation from a flow matching teacher instead of one-point velocity MFT (ablation) | MF-RAE on RAE latents · ImageNet 256 1-step | FID: 2.81 -> 2.03 | Improved |
| use MFD instead of MFT on SD-VAE latents with guidance (ablation) | vanilla MF with SD-VAE · ImageNet 256 1-step | FID: 3.38 -> 3.15 | Improved |
| initialize MF from Consistency Mid-Training weights instead of random or flow matching init (ablation) | MF on RAE latents · ImageNet 256/512 training stability | training stability | Improved |
| add MFT bootstrapping stage after MFD (ablation) | MF-RAE · ImageNet 512 1-step | FID: 3.95 -> 3.23 | Improved |
| add MFT bootstrapping stage after MFD at 256 resolution (ablation) | MF-RAE · ImageNet 256 | FID | No clear effect |
| replace exact JVP with finite difference, dt between 1e-3 and 1e-2 (ablation) | MF-RAE · ImageNet 256 1-step | FID: JVP 1.96 vs 2.03 at dt 5e-3 | No clear effect |
| use too small finite-difference step dt=1e-4 (ablation) | MF-RAE · ImageNet 256 1-step | FID: 2.03 -> 5.63 | Hurt |
| remove classifier-free guidance by distilling an unguided class-conditional teacher | MF-RAE · ImageNet 256 | hyperparameter complexity | Improved |
| replace SD-VAE decoder with RAE decoder | MeanFlow sampling · ImageNet 256 1-step | sampling GFLOPS: 424 -> 263 | Improved |
| use MF-RAE pipeline instead of training vanilla MF from scratch | ImageNet 256 training · ImageNet 256 | training cost (H100 GPU-days): over 600 -> about 100 | Improved |
| naive MF training on RAE latents without CMT (ablation) | MF on RAE latents · S and B scales | 1-step FID before divergence: best FID still above 20 | Hurt |
Closest papers by meaning
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.88)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.88)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.87)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.87)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.87)
- Improved Baselines with Representation Autoencoders (NeurIPS 2026 · 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.
