📢📢📢 Introducing Spaghetti Splatting
3DGS is a convolution of a Dirac with a Gaussian Kernel. We provide a strict generalization of 3DGS, extending its Dirac mass to lines and curvilinear structures
project page: https://t.co/oGPJAuBE4O
arXiv: https://t.co/2Ql1Edi9JJ
🔊 on!
🔦 Holy moly. Optimized the splatlight renderer to widen from 25° to 42°. Even more performant, locked at 90 FPS on Meta Quest 3, and way more immersive.
The more I play with @theworldlabs Atlas, the more conviction I have we live in someone else’s prompt.
Looking for 3D Gaussian splat content for Blender? Install the open source SuperSplat Blender Extension from @playcanvas.
🔍 Search downloadable splats on SuperSplat
📋 Copy/paste the URL to Blender
🪄 Splat appears in Blender's viewport
https://t.co/a89nNrvBO1
@ST0NEHENGE Am no bein funny (or biased. They're guid folks the Scots, but ah wouldnae trust any o' they experts, ken? There's millions ae other stoned experts up there🤣
@PattayaUnf I lived in BKK for four years, got the t-shirt. I thank the bar owner in Pattaya who wrote the (free) book Private Dancer. That should be compulsory reading for any noobs heading down there. It explains how they do it and theres Thai recipes in it 😂
abusing splats pt 2: responsive web interfaces
so it turns out gaussian splats can render crisp text and buttons just fine, not to mention video (decoded in realtime). they can also handle viewport changes and hover states, and there's a real benefit to driving splats with markup: you get splat field *scope* for free. knowing which splats correspond to which element from the outset means you get to work with them semantically instead of just as a million blurry patches of color
that said, the splat counts get wild fast, with multiple millions for in-viewport content on a relatively simple page. adaptive splat density makes a meaningful difference for counts at desktop scale, letting you cut them by 20%+ without losing text detail
and while this experiment does NOT cheat by rendering full webpage paints to splat space, it does still rely on the browser to give us text bounding rects, css geometry + layout rules, and rich media one frame at a time. it also worked my gpu pretty hard just do what standard web rendering already does well, but... it did do it!
now I'm curious: given all the effort it takes to get to this fairly pedestrian outcome, are there any interesting experiences splats unlock? tbd, stay tuned
Iv added the SHARP gaussian splat monocular model to the Solv3R setup.
allowing me to use the positional data from Solv, and the splat from SHARP to create a more complete gaussian splat. a quick few thousand step training on the two models aligns their points in space to achieve this!
The results are not good what i would call good. but it does work!
EDGAR sharpens a Gaussian head avatar with identity-guided diffusion, then bakes the fixes back into the 3D model. Refinement runs once so the original renderer stays as fast as before; training and inference code are public.
https://t.co/6noexZuxiK