The paper introduces RAWATTA, a training‑free technique that accelerates multi‑view 3D transformers by approximating global attention in 8‑bit precision, selectively retaining 16‑bit precision only where necessary. It identifies three functional roles of transformer layers—averaging, matching, and anchoring—and applies different approximations accordingly, including a faster variant that further simplifies averaging and anchoring layers. On a 12 GB GPU with 50 images, RAWATTA speeds up VGGT while losing at most one point in pose accuracy, and both variants can handle 200 images, surpassing the released code’s memory limits.
By Muhammed Izzet Sezer, Sevval Belkis Dikkaya
BlazeVGGT is a training‑free, plug‑and‑play module that speeds up the VGGT visual geometry model by sparsifying cross‑view attention. It selects a compact set of complementary views using a greedy determinantal point process, preserving dense per‑frame predictions while eliminating redundant computations. On 2,000‑frame sequences, BlazeVGGT achieves roughly 13× faster inference and improves Chamfer Distance from 0.41 to 0.32, demonstrating a strong accuracy–efficiency trade‑off for point‑cloud reconstruction and camera‑pose estimation.
By Shujun Yao, Junhui Lin, Congying Han, Tiande Guo, Ziwen Liu, Bonan Li
arXiv:2605.16981v3 Announce Type: replace
Abstract: Streaming 3D reconstruction under a strict constant-memory budget hinges on how the recurrent state is updated as the stream evolves. We profile TT...
By Kejun Ren, Lei Jin, Tianxin Huang, Lianming Xu, Li Wang
arXiv:2609.23733v1 Announce Type: new
Abstract: Feed-forward visual geometry models such as the Visual Geometry Grounded Transformer (VGGT) have recently enabled direct 3D reconstruction from multi-v...
By Abteen Arab, Guile Wu, Chengjie Huang, Dongfeng Bai
The study investigates how long‑video language models decide which frames to keep, compress, and reuse, testing each decision in isolation across six selection rules, three benchmarks, and two answering models. It finds that selecting frames based on queries yields the biggest performance boost, that halving spatial resolution costs little, and that reallocating saved tokens to more compressed frames can further improve accuracy. The work also highlights the importance of a unified evaluation harness to avoid misleading comparisons.
By Prakhar Khatri
GPA‑VGGT is a training‑free acceleration technique for the VGGT feed‑forward geometry model. It replaces VGGT’s global attention in early blocks with frame‑wise attention that uses a shared cross‑frame key/value set, and in later blocks it merges only keys and values while preserving queries and restoring attention mass. The method achieves speedups comparable to existing efficiency approaches while maintaining over 99% of VGGT’s pose accuracy on multi‑view geometry tasks.
By Muyeol Sung, Jangwon Kim, Jongwoo Lim
arXiv:2510. 09608v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) could power real-time assistants and autonomous agents, but they face a critical challenge: understanding near-infinite video streams without escalating latency and memory usage.
By Ruyi Xu, Guangxuan Xiao, Yukang Chen, Liuning He, Yao Lu, Song Han
Detectors for AI-generated video are evaluated offline. A clip is decoded to pixels and scored once, increasingly by a large vision-language model.
CoDecide is a retrospective analytics system that selectively decodes video frames only when new content appears, using compressed‑bitstream signals to identify novelty. It trains a lightweight compressed‑domain student model with a label‑free novelty signal from a frozen DINOv2 teacher, enabling it to generalize across camera deployments without per‑camera retraining. In experiments on VIRAT, MEVA, and ShanghaiTech, CoDecide outperforms prior compressed‑domain selectors and a dense per‑frame detector, achieving 15 to 24× faster processing while recovering more moving‑object detections under the same decoding budget.
By Shristy Shah, Thomas B. Moeslund, Kamal Nasrollahi
arXiv:2609.15810v1 Announce Type: new
Abstract: Diffusion Transformers deliver state-of-the-art video generation, but their long spatiotemporal sequences make attention the dominant deployment cost,...
By Xingyang Li, Dongyun Zou, Shining Zhang, Jiacheng Chen, Haocheng Xi, Lvmin Zhang, Jun-Yan Zhu, Song Han, Zhekai Zhang, Yujun Lin, Muyang Li
arXiv:2606. 01790v1 Announce Type: cross Abstract: Vision-language-model-based graphical user interface (GUI) agents have shown broad automation capabilities, yet deployment is bottlenecked by a key-value (KV) cache that grows linearly with interaction steps.
By Yuhang Han, Wenzheng Yang, Yujie Chen, Xiangqi Jin, Yaojie Zhang, Siteng Huang, Linfeng Zhang
The paper introduces FrameBlend, a plug‑and‑play technique for streaming video understanding that selectively recomputes key‑value (KV) caches at query time. By re‑prefilling only query‑relevant video tokens, FrameBlend allows shorter attention windows during idle prefill, cutting prefill cost while boosting accuracy. Experiments on three SVU frameworks and five benchmarks show up to 3.9‑point accuracy gains and significant reductions in FLOPs and processing time.
By Junhyeok Lee, Sangyoon Lee, Jaeho Lee