We put FLUX 3's native layout control to the test on OverLayBench, our benchmark for layout-to-image generation with dense overlaps. The results blew us away. 🔥 On our hardest split, @bfl_ai's FLUX 3 Image Klein hits 54.94 O-mIoU vs. 37.82 for the previous best (+17.12).
How we ran it: the boxes go in as FLUX 3's JSON layout, no prompt upsampling, 1024×1024, one seed, scored with the standard OverLayBench toolkit (Qwen2.5-VL-32B judge). Want your model on the leaderboard? Reach out to @LibnLee.
On the hardest split it leads all 19 previous methods on the leaderboard, and it beats every one of them on 19 of the 21 leaderboard columns across the three splits.
The harder the scene, the bigger the lead. O-mIoU margin over the best previous method: +11.1 on Simple, +15.1 on Regular, +17.1 on Complex. Its instance success rate barely moves (97.61 → 97.15 → 94.88), while the best previous drops from 88.80 to 77.87.
Explore the updated leaderboard, benchmark, and evaluation tools: mlpc-ucsd.github.io/OverLayB… For context: from GLIGEN (Jan 2023) to Laytrol (Nov 2025), the best O-mIoU on our hardest split rose +13.97 points, to 37.82. FLUX 3 Image Klein adds +17.12 in one step.
Huge congrats to the Black Forest Labs team! It's exciting to see the frontier move this much, and we'd love to see more models take on the challenge.
OverLayBench targets what models struggle with: crowded scenes, heavily overlapping boxes, and similar-looking subjects that have to stay distinct. O-mIoU scores only the overlapping regions. FLUX 3 Image Klein also gets 74.25 mIoU and a 94.88% instance success rate on Complex.
Bingnan Li retweeted
Long videos should remember what appeared, not when. 🎬
Introducing RECAP-Forcing, a training-free memory for long video generation. It stores each new subject, object and scene once, at first sight, and keeps it, so minute-long videos stay consistent and keep moving. 🧵
Bingnan Li retweeted
🚀 Excited to share our new work - LIFT: Layout-In-Future Video Generation under Large Viewpoint Change via On-Policy Self-Distillation!
🌐 Project page: jsxzs.github.io/LIFT/
📄 Paper: huggingface.co/papers/2609.3…
LIFT enables controllable image-to-video generation with:
📷 Large viewpoint change
📦 Future-view layout control
🔄 On-policy self-distillation (OPSD)
Instead of requiring dense per-frame layouts, LIFT only needs last-frame object layouts to control what appears and where in the future view.
Bingnan Li retweeted
AI can draw a dog and a cat side by side, no problem. But ask it for “one golden retriever half-hidden behind another,” and you’ll get a four-legged blob. Why? Because most layout-to-image models are tested on scenes where objects don’t touch. Real photos don’t work that way.
In our #NeurIPS2025 paper, OverLayBench: A Benchmark for Layout-to-Image Generation with Dense Overlaps (mlpc-ucsd.github.io/OverLayB…), @jianwen_xie and @zx1239856 introduce a new way to address it. The OverLayScore metric measures layout difficulty by combining box overlap and object similarity that you can use to build a benchmark full of crowded, overlapping scenes.
The fix? Train models with amodal masks so they learn the full shape of an object even when parts of it are hidden. This improves overlap-region accuracy by 16% with zero extra inference cost.
#ComputerVision
Bingnan Li retweeted
🚀 Introducing OverLayBench – A Benchmark for Layout-to-Image Generation with Dense Overlaps! [NeurIPS 2025]
📜 Paper: arxiv.org/abs/2509.19282
🌐 Project: mlpc-ucsd.github.io/OverLayB…
💻 Code: github.com/mlpc-ucsd/OverLay…
🔍 The Problem:
Current Layout-to-Image (L2I) models collapse under heavy bounding-box overlaps: objects blend together, distort, or misalign. Existing benchmarks are biased toward simple cases, leaving this critical challenge underexplored.
🤔 Why It Matters:
Dense overlaps are everywhere in the real world—from crowded streets 🚦 to busy dining tables 🍽️. Evaluating and improving L2I under such conditions is key for robust controllable generation.
💡 Our Solution:
✅ OverLayScore – a new metric capturing both spatial IoU and semantic similarity to quantify overlap complexity.
✅ OverLayBench – a benchmark with balanced difficulty splits (Simple / Regular / Complex), high-quality captions, instance-level grounding, and relationship annotations.
✅ CreatiLayout-AM – a new baseline using amodal mask supervision, significantly improving generation quality under occlusion (+15% O-mIoU).
Check out more results in the thread below. 👇 #NeurIPS2025
Bingnan Li retweeted
We’re thrilled that our lab’s work on “Deeply-Supervised Nets” has received the Test-of-Time Award at AISTATS 2025! 🏆 This prestigious award honors papers published 10 years ago that have had a lasting and significant impact on the field of artificial intelligence and statistical modeling.
DSN combats the vanishing and exploding gradient problem when training deep neural networks by having each individual layer directly access the supervised labeling information. This significantly facilitates the training of deep networks.
Congratulations to all authors of this work: Chen-Yu Lee (@chl260), Saining Xie (@sainingxie), Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu.
Paper: arxiv.org/abs/1409.5185
Award slides: chl260.github.io/PDF/dsn-ais…
Official announcement: xitter.cf/aistats_conf/status/19…
@UCSanDiego @UCSDCogSci @UCSDJacobs @aistats_conf
The #AISTATS 2025 Test of Time Award goes to ... 🥁 ...
Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, Zhuowen Tu, for "Deeply Supervised Nets"! Congratulations!
