Benjamin Henriksson

@noquierocoima @ben_sdl Kind of? FlashSplat figures out how much each gaussian contributes to any pixel, then uses many 2D segmentations to find which gaussians match a segmented object. You get "scores" for how closely each gaussian links to a concept, which I just thresholded in the viewer.

@BenjaminEmHe @ben_sdl in every splat have segment info?

Semantically segmented Gaussian Splat.

@pablo_troyse @lichtfeldstudio The source images were actually taken from a camera rig mounted on an airplane, which was somewhat challenging to work with due to viewpoint sparsity and the extreme resolution of the images (roughly 20k by 14k).

@BenjaminEmHe @lichtfeldstudio What drone did you use?

~800-hectare Gaussian splat, trained with @lichtfeldstudio.

@ben_sdl I used SAM 3 with text prompts on the source imagery and then projected it into 3D with FlashSplat!

@BenjaminEmHe Super cool, how did you classify the objects ?

@bilawalsidhu @alexanderchen Also possible to do volumetrically! x.com/BenjaminEmHe/s…

Semantically annotating 3D gaussian splats on the fly using gemini 3.1 + sparkjs 1. Load any 3D scene and hit scan 2. Get 2D detections from VLM 3. Cluster outputs & project into 3D world space 4. Save as a persistent 3D semantic layer Inspired by @alexanderchen's experiments with gemini visual intelligence. Just had to try to lift it from 2D to 3D!

@nylanderjens Tack!

@BenjaminEmHe Om priset skiljer sig markant ska den upphandlade organisationen begära in en förklaring. Då förklarar man exakt hur man beräknat och vad som gör att man kan leverera för priset, vi har sådan standardrutin som inleds med...

En myndighet har betalat 6,3 MKR/år för enkel data. Mitt AI-bolag lämnar in ett anbud på 0,9 MKR vs 7,5 MKR/år. Konkurrenten överprövar och tycker det är för billigt m.m. AI kommer plocka bort allt "fluff" som finns. AI-startups ska in i upphandlingar - fråga gärna om råd.

@XRarchitect 👀

Pumped to be featured in the latest Midjourney Issue 36 I’ve been using it as an iteration system to explore inputs for persistent 3D creation with World Labs across the web and mixed reality. Excited about where this is all heading 🌎

@naribubu Any recommendations for ICP packages? Had some trouble aligning LIDAR and SfM point clouds before (especially with drone images and ground LIDAR) :/

Livox mid 360 + GLIMでSLAMした点群にポスト処理で色付けしてみた。 手順： ①OSMO360で同じ場所を撮影 ②12方向reframeしてColmapでSfM ③SfMの点群とLiDAR点群をICPで合わせて変換行列を得る ④Colmapのimages.txtを上記変換行列でLiDAR座標に合わせる ⑤LiDAR点群をカメラの画像座標系に透視投影変換する ⑥LiDAR点群に色を付ける（一番近いカメラを選択する） ※LiDARとOSMOは撮影日が異なり移動した車は色付けズレてます。また地面は自身が映りこんでる(-_-;)

@martin_casado @sparkjsdev Thank you!

@BenjaminEmHe @sparkjsdev Here is the branch. Although it's incomplete and needs to be cleaned up. But it has the LoD and streaming code. github.com/sparkjsdev/spa…

Insane LoD/streaming test of experimental @sparkjsdev branch. 100 scenes with ~150m splats. 16million local splat budget for streaming, 2million visible budget for LoD. Once this stabilizes I think we can support persistent splat worlds of arbitrary size ...

@martin_casado @sparkjsdev Absolutely incredible. Is there a link to this branch?

@BenjaminEmHe @sparkjsdev No, browser. It can do 100 fps+ on my laptop. Around 47fps on my quest 3.

@DSkaale Interesting, might it work with image segmentation models? e.g. SAM3, etc.

🎨 WIP: My Gaussian Splatting editor now has a revolutionary color-coded copy/paste workflow. Addition to box tools, I'm using dual GPU selection buffers visualized as colored overlays. You just paint what you want and paint where it goes. Technical Highlights: • Dual-buffer selection: Orange for source, Blue for target regions. • Auto-alignment: Captures camera orientation to compute target-to-source quaternion deltas. • Occlusion culling: Two-pass GPU depth test using InterlockedMin to filter for front-most splats. • Real-time refinement: Specialized compute kernels for incremental transforms with zero Euler drift. • Performance: Sub-millisecond latency for 100k+ splats via massive GPU parallelism. Still in development but already changing how I edit splats. #GaussianSplatting #Unity3D #ComputeShaders #GameDev #WIP #GameDev

@nearcyan Yeah, unfortunately the Quest 3 struggles performance-wise with viewing splats through WebXR etc. (although there is some form of experimental native support?). Would be nice to see more optimizations for native viewing.

Meta and Apple put a ton of resources into this, and it doesn't seem to get much attention yet because XR hardware is still rare as soon as real XR glasses come out (...so never), the quality will be pretty amazing. i'm already amazed with SOTA if you feed it 100s of images

gaussian splatting quietly yet rapidly improving

Gaussian splat from Kista, Stockholm.

@SadlyItsBradley Very curious, is this being viewed with an Apple Vision Pro?

I tried out that newly released Apple “single image to gaussian splat” model today It’s actually incredible. Unlike the current spatial scenes feature, you can actually WALK INTO the memory you captured on your phone It only took about 10 seconds on my MacBook Pro to generate

Messing around with Flux.2 for upscaling OOD viewpoints in a Gaussian splat, works surprisingly well running locally on an RTX 5090.

@cixliv @NimaZeighami @the_carlosdp What headset is being used? 👀

POV you turn a small Airbnb kitchen into a humanoid robot testing room before a NYC robot fight. The Airbnb just left me a positive review, I can post this picture now.

@janusch_patas Would love to get Skyfall GS running, but I can't seem to find any pre-trained checkpoints. Can they be found anywhere, or does it have to be trained from scratch?

3DGS, satellite imagery, and drones are game changers for disaster recovery and modern warfare. Progress in real-time 3D reconstruction will lift the fog of war. Drones are cheap and ubiquitous, and with Gaussian Splatting, anyone can build near-instant spatial awareness. This can even happen in a feed-forward way, similar to how Tesla approaches autonomous vision. The first LichtFeld Studio bounty already demonstrated that seven scenes can be trained in about twenty minutes on an RTX 4090, even with a classical optimization-based approach. Combine this with feature embeddings that let you prompt directly within the scene for objects like humans, tanks, or buildings. Almost every week, new research papers are published doing exactly this. Projects like Skyfall GS already turn satellite images into explorable 3D urban environments using diffusion models with real-time rendering performance. Companies are beginning to experiment with decentralized mapping, where different actors contribute their own 3D data that merges into one coherent world model. In disaster zones, this means instant 3D situational maps. Collapsed buildings, flooded streets, and blocked roads can be reconstructed and shared within minutes to guide rescue teams. In warfare, it becomes a massive intelligence amplifier, combining drone and satellite imagery into live 3D maps of terrain, movement, and infrastructure faster than any traditional reconnaissance could. The next step is not just seeing the world but reconstructing it in real time.