Research map / Image generation / Autoregressive and latent generation
MIMFlow: Integrating Masked Image Modeling with Normalizing Flows for End-to-End Image Generation
Yang Chen, Xiaowei Xu, Shuai Wang, Xinwen Zhang, Qiushi Guo, Tiezheng Ge, Limin Wang
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 masking (ratio 0.4-0.6) to encoder input (ablation) | MIMFlow · ImageNet 256x256 (50 ep.) | gFID: 29.0 -> 12.82 | Improved |
| add masking (ratio 0.4-0.6) to encoder input (ablation) | MIMFlow · ImageNet linear probing | top-1 accuracy: 56.6 -> 71.3 | Improved |
| use mixed mask/no-mask strategy or low mask ratios (ablation) | MIMFlow · ImageNet 256x256 (50 ep.) | gFID: 12.82 -> 26.98 (mix); 24.47 (0.2-0.4) | Hurt |
| add masking bottleneck in progressive ablation (ablation) | MIMFlow · ImageNet 256x256 (no CFG) | gFID: 18.71 -> 10.14 | Improved |
| add DINO+CLIP auxiliary supervision (ablation) | MIMFlow · ImageNet 256x256 (no CFG) | gFID: 19.52 -> 18.71 | No clear effect |
| use learnable tokens while removing GAN loss (ablation) | MIMFlow baseline · ImageNet 256x256 (no CFG) | gFID: 11.24 -> 19.52 | Hurt |
| add GAN fine-tuning of decoder (ablation) | MIMFlow · ImageNet 256x256 (no CFG) | gFID: 10.14 -> 6.15 | Improved |
| add HOG as auxiliary target (ablation) | MIMFlow · ImageNet 256x256 | gFID: 12.46 -> 20.77 (CLIP+HOG) | Hurt |
| combine DINO and CLIP auxiliary targets instead of DINO only (ablation) | MIMFlow · ImageNet 256x256 | gFID: 12.89 -> 12.46 | Improved |
| increase latent tokens from 128 to 192 (ablation) | MIMFlow · ImageNet 256x256 | gFID: 12.46 -> 30.42 | Hurt |
| use latent noise sigma 0.5 instead of 0.3 (ablation) | MIMFlow · ImageNet 256x256 | gFID: 12.82 -> 17.69 | Hurt |
| use latent noise sigma 0.2 instead of 0.3 (ablation) | MIMFlow · ImageNet 256x256 | gFID: 12.82 -> 14.30 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- STARFlow-XXL — STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025) · better on 0 of 1 shared result
- 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
- MaCow: Masked Convolutional Generative Flow (NeurIPS 2019 · similarity 0.89)
- STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis (NeurIPS 2025 · similarity 0.88)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.88)
- JanusFlow: Harmonizing Autoregression and Rectified Flow for Unified Multimodal Understanding and Generation (CVPR 2025 · similarity 0.87)
- Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis (ICLR 2025 · similarity 0.87)
- Fixed-Point Masked Generative Modeling (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.
