🌈FlowWAM: flow->action prediction🌈

👉FlowWAM is a novel dual-stream diffusion framework that adopts optical flow as a unified, video-native action representation. Repo under Apache💙

👉Review https://t.ly/FmutT
👉Paper https://arxiv.org/abs/2607.13017
👉Project https://flow-wam.github.io/
👉Repo github.com/YixiangChen515/FlowWAM

🏯SOTA Music-to-Dance Gen🏯

👉The Tongyi Lab unveils Wan-Dancer, a novel stable minute-scale synthesis at 720p/30fps across five dance genres. Impressive results, new SOTA on long clip by a large margin. Repo under Apache 2.0💙

👉Review https://t.ly/AKY5j
👉Paper https://lnkd.in/d_xA7dwb
👉Project https://lnkd.in/dzfnw2h4
👉Repo https://lnkd.in/d-Zj_cTf

👉Not a render. Not a concept. This is GENE.01 by Generative Bionics, the Italians coolest scaleup strikes back: in just six months, they turned GENE.01 into a fully functional humanoid platform that can walk, sense and interact.

👉Full-body multimodal skin perceives touch, proximity, force and temperature, bringing Physical AI closer to safe and natural collaboration with people.

👉More: https://t.ly/F3I3A

💢Unified Video Dense Prediction💢

👉UniD predicts: depth, surface normals, semantic segmentation, boundaries, human parts, albedo, shading, and materials. Code TBR💙

👉Review https://t.ly/oo7et
👉Paper https://arxiv.org/pdf/2607.21592
👉Project https://unid-video.github.io/
👉Repo https://github.com/YihongSun/UniD

🍿 Dawn of Generative Cinematography 🍿

🟩 #TheOdyssey by Christopher Nolan was shot entirely on IMAX 70mm. It feels almost romantic: massive cameras, film stock, premium lenses, and an obsessive pursuit of the highest possible quality at the moment of capture.

👉 Meanwhile, #AI research is heading in the exact opposite direction.

🟩 A pre-print paper released today, "Camera Anything", demonstrates something that sounded like science fiction just a few years ago: you film a scene once... and then you can virtually reposition the camera anywhere.

👉More https://t.ly/Kd7RV
👉Paper arxiv.org/pdf/2607.24591
👉Project yixuanli98.github.io/cameraanything/
👉Repo github.com/yixuanli98/CameraAnything

🐠Dual-branch ID-Tracking🐠

👉TIDE: tracking dense, homogeneous targets, providing a scalable dual-branch design to accommodate diverse hardware constraints. MIT license💙

👉Review https://t.ly/WEDeY
👉Paper https://arxiv.org/pdf/2607.26412
👉Project https://vranlee.github.io/TIDE/
👉Repo https://github.com/vranlee/TIDE

🔥Unified Points n' Lines🔥

👉ETH (+Microsoft Spatial AI Lab) unveils a novel feature extractor that jointly extracts keypoints, lines, and feature descriptors within a single lightweight net. SOTA in line detection can be achieved by adding only three convolutional layers to existing point extractor. Repo under Apache💙

👉Review https://lnkd.in/p/eW8j5JZj
👉Paper https://arxiv.org/pdf/2608.19894
👉Repo https://github.com/francois141/upal

🐆 Anyone in 4D is out 🐆

👉4DAnyone turns a casual monocular video into multi-view videos, enabling downstream 4DGS reconstruction. Full repo under Apache 2.0💙

👉Review https://lnkd.in/p/ec4dzGvb
👉Paper https://arxiv.org/pdf/2608.20335
👉Project https://4danyone.github.io
👉Repo github.com/ant-research/4DAnyone

🍋‍🟩Remesh-Aware Mesh Deformation🍋‍🟩

👉RADmesh is a novel generative deformation technique enhanced by remeshing. Given a text prompt, it deforms and remeshes a mesh region to form new geometric features. Repo MIT💙

👉Review https://lnkd.in/p/eK6FZv9c
👉Paper https://arxiv.org/pdf/2608.17182
👉Project https://threedle.github.io/radmesh/
👉Repo https://github.com/threedle/radmesh/

👻Emerging Objs from Motion👻

👉Motion boundaries provide a strong signal for object-level grouping and can be used to derive pseudo-instance supervision. Suitable for: mono-depth, 3D object detection, 3D occupancy, and end-to-end planning. Repo under Apache 2.0💙

👉Review https://lnkd.in/p/eezZrSJE
👉Paper https://arxiv.org/pdf/2609.04348
👉Project https://tj12342.github.io/object-concepts-from-motion/
👉Repo https://github.com/TJ12342/object-concepts-from-motion/tree/main

🏀McByte++ tracking-by-detection🏀

👉McByte++ is the newer extension of McByte that advances training-free sports MOT toward long-term ID tracking, while simultaneously improving efficiency and runtime performance. Repo under Apache 2.0💙

👉Review https://lnkd.in/p/e4-diVJS
👉Paper https://lnkd.in/e_Vxky-b
👉Repo https://lnkd.in/e8SeCYmk

🔥 RelateAnything is gold! 🔥

👉RelateAnything is a 53M-parameter relation model that takes an image and a set of regions from any source and returns scored relations over a predicate vocabulary supplied at inference as a list of strings. Impressive results. Repo under Apache 2.0💙

👉Review https://lnkd.in/p/etAcdFM3
👉Paper https://arxiv.org/pdf/2609.12552
👉Repo https://github.com/Maelic/RelateAnything
👉Project https://maelic.github.io/RelateAnythingProject/