Research map / Image generation / Autoregressive and latent generation
RecTok: Reconstruction Distillation along Rectified Flow
Qingyu Shi, Size Wu, Jinbin Bai, Kaidong Yu, Yujing Wang, Yunhai Tong, Xiangtai Li, Xuelong Li
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 flow semantic distillation (FSD) on x_t (ablation) | RecTok · ImageNet 256 ablation | gFID: 3.35 -> 2.27 | Improved |
| add flow semantic distillation to baseline (ablation) | RecTok · ImageNet 256 ablation | gFID: 12.07 -> 2.52 | Improved |
| add reconstruction distillation (RAD) after FSD (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.52 -> 2.27 | Improved |
| add decoder finetuning (ablation) | RecTok · ImageNet 256 ablation | rFID: 0.65 -> 0.48 | Improved |
| add reconstruction alignment to alignment-only (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.52 -> 2.27 | Improved |
| add alignment to reconstruction-only (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.97 -> 2.27 | Improved |
| increase latent dimension from 16 to 128 (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.75 -> 2.27 | Improved |
| use shift noise schedule instead of uniform (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.50 -> 2.27 | Mixed |
| use random encoder initialization instead of VFM initialization (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.37 -> 2.27 | Improved |
| use lightweight 1.5M transformer semantic decoder instead of 10M (ablation) | RecTok · ImageNet 256 ablation | gFID: 3.13 -> 2.27 | Improved |
| set semantic loss weight to 1 instead of 0.5 (ablation) | RecTok · ImageNet 256 ablation | gFID: 2.78 -> 2.27 | Improved |
| use two VFMs instead of one for semantic learning (ablation) | RecTok · ImageNet 256 ablation (dim 128) | gFID: 2.27 -> 2.51 | Hurt |
Closest papers by meaning
- Aligning Visual Foundation Encoders to Tokenizers for Diffusion Models (ICLR 2026 · similarity 0.90)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.89)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.88)
- Cross Flow: One-Step Generation Across Latent and Pixel Spaces (NeurIPS 2026 · similarity 0.88)
- Structured Coupling for Flow Matching (NeurIPS 2026 · 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.
