Research map / Image generation / Autoregressive and latent generation
Flow to the Mode: Mode-Seeking Diffusion Autoencoders for State-of-the-Art Image Tokenization
Kyle Sargent, Kyle Hsu, Justin Johnson, Li Fei-Fei, Jiajun Wu
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 mode-seeking post-training (Stage 1B) (ablation) | FlowMo-Lo · ImageNet-1K tokenization | rFID: 1.10 -> 0.95 | Improved |
| add mode-seeking post-training (Stage 1B) (ablation) | FlowMo-Hi · ImageNet-1K tokenization | rFID: 0.73 -> 0.56 | Improved |
| add mode-seeking post-training (Stage 1B) (ablation) | FlowMo-Lo · ImageNet-1K tokenization | PSNR: 21.38 -> 22.07 | Improved |
| add mode-seeking post-training (Stage 1B) (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 2.02 -> 1.21 | Improved |
| add mode-seeking post-training (Stage 1B) (ablation) | FlowMo-Continuous · ImageNet-1K tokenization | rFID: 0.74 -> 0.65 | Improved |
| add perceptual loss on 1-step prediction (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 13.86 -> 2.87 | Improved |
| train encoder end-to-end rather than with MSE (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 3.82 -> 2.87 | Improved |
| use thick-tailed logit-normal noise schedule (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | PSNR: 16.45 -> 20.71 | Improved |
| use halved patch size (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 6.39 -> 2.87 | Improved |
| use LFQ instead of FSQ (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 3.14 -> 2.87 | Mixed |
| use shifted sampler (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 3.42 -> 2.87 | Improved |
| add guidance in a limited interval (ablation) | FlowMo (fewer params) · ImageNet-1K tokenization | rFID: 3.28 -> 2.87 | Improved |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.90)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.90)
- End-to-End Training for Unified Tokenization and Latent Denoising (NeurIPS 2026 · similarity 0.89)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.89)
- Epsilon-VAE: Denoising as Visual Decoding (ICML 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.
