Research map / Probabilistic and scientific modeling / Variational and generative sampling
Generative Flows With Invertible Attentions
Rhea Sanjay Sukthanker, Zhiwu Huang, Suryansh Kumar, Radu Timofte, Luc Van Gool
This paper tried a common change: More attention heads
Use more than 5 iTrans attention heads · CIFAR10 · bits/dim No clear effect
What 107 accepted papers report for this change
- 62 improved
- 13 no clear effect
- 13 mixed
- 19 hurt
- Universal Approximation with Softmax Attention ICML 2026 · increase the number of attention heads H · synthetic truncated linear model (d=10, n=50) · test MSEImproved
- TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning ICML 2026 · raise number of heads from 10 to 14 · ScienceQA SIG detection · F1: 89.74 -> 77.09Hurt
- Efficient Neural Controlled Differential Equations via Attentive Kernel Smoothing ICML 2026 · increase number of attention heads from 1 to 8 · UEA classification (3 datasets) · error rate: error rates reduced as M increases from 1 to 8; plateau around M = 4 for GPImproved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
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 iMap attention to mARFlow | mARFlow · MNIST | bits/dim: 0.56 -> 0.39 | Improved |
| Add iMap attention to mARFlow | mARFlow · CIFAR10 | bits/dim: 3.24 -> 3.216 | Improved |
| Add iMap attention to mARFlow | mARFlow · CIFAR10 | FID: 41.9 -> 33.6 | Improved |
| Add iTrans attention to mARFlow | mARFlow · CIFAR10 | bits/dim: 3.24 -> 3.217 | Improved |
| Add iMap attention to SRFlow | SRFlow · CelebA 8× SR | SSIM: 0.73 -> 0.74 | Improved |
| Add iTrans attention to SRFlow | SRFlow · CelebA 8× SR | LPIPS: 0.17 -> 0.16 | Improved |
| Add iMap attention to Full-Glow | Full-Glow · Cityscapes label→photo | conditional bits/dim: 2.345 -> 2.310 | Improved |
| Add iTrans attention to Full-Glow | Full-Glow · Cityscapes label→photo | conditional bits/dim: 2.345 -> 2.314 | Improved |
| Replace spatial-channel checkerboard mask with spatial-only mask (ablation) | AttnFlow-iMap · MNIST | bits/dim: 0.39 -> 0.99 | Hurt |
| Replace spatial-channel checkerboard mask with channel-wise mask (ablation) | AttnFlow-iMap · MNIST | bits/dim: 0.39 -> 0.75 | Hurt |
| Replace spatial-channel checkerboard mask with random binary mask (ablation) | AttnFlow-iMap · MNIST | bits/dim: 0.39 -> 0.50 | Hurt |
| Place attention after permutation or after coupling rather than earlier in the flow step (ablation) | AttnFlow · CIFAR10 | bits/dim | Improved |
Closest papers by meaning
- Attention Distillation: A Unified Approach to Visual Characteristics Transfer (CVPR 2025 · similarity 0.86)
- Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (ICML 2024 · similarity 0.86)
- Expanding Flow Maps (NeurIPS 2026 · similarity 0.86)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · similarity 0.86)
- CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up (NeurIPS 2025 · similarity 0.86)
- LaTtE-Flow: Layerwise Timestep-Expert Flow-based Transformer (ICML 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.
