Ruilong Li
@ruilong_li
Followers
1,440
Following
278
Media
15
Statuses
129
PhD student at @berkeley_ai | Alumni of @NVIDIAAI , @GoogleAI , @RealityLabs , @USC_ICT , @Tsinghua_Uni
Berkeley, CA
Joined August 2019
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Puerto Rico
• 573697 Tweets
더보이즈
• 53821 Tweets
Mustafa Kemal Atatürk
• 38992 Tweets
ねないこだれだ
• 34610 Tweets
Efendiler
• 34092 Tweets
$ICE
• 30378 Tweets
Duterte
• 27144 Tweets
Happy New Week
• 23356 Tweets
CDTV
• 21485 Tweets
#リリリエラ
• 20256 Tweets
KabinetPRABOWO GIBRANkompak
• 15639 Tweets
Langsung GASSS Kerja CEPATTT
• 14895 Tweets
Day3
• 13965 Tweets
レッドカーペット
• 13721 Tweets
#GOAT_80Mviews
• 12963 Tweets
पप्पू यादव
• 12208 Tweets
国民玉木雄一郎代表
• 11372 Tweets
Pinned Tweet
Excited to announce 🚀gsplat v1.0🚀: a ⏩efficient⏩ CUDA backend for 3D Gaussian Splatting!
A drop-in replacement of the official impl. with:
- Up to 2x faster training;
- Up to 4x less GPU memory;
- Render millions of GSs in real-time;
- And more;
10
109
509
Excited to release 🚀NeRF-XL🚀, a principled method to distribute your NeRF's parameters to multiple GPUs!
It is heuristics-free and mathematically equiv. with training your NeRF on a single (large) GPU.
Let's scale it up!🌎
17
81
324
Forget to show off this fish-eye results from gsplat's impl.! Super cool right?
Note this combines other best techs (3DGS-MCMC, bilateral grid) because everything is modularized in gsplat!
🌟gsplat==1.4.0🌟 () is now released!
- Supports 2DGS! (Kudos to
@WeijiaZeng1
)
- Supports Fish-eye camera! (Kudos to
@jefequien
)
- Prebuilt wheels for Windows! (Kudos to Martin de La Gorce)
2
26
232
7
39
315
Excited to release ⏩nerfacc⏩: A NeRF acceleration toolbox for fast differentiable volumetric rendering!!! Plug-and-play for most of NeRFs!! And it is with pure Python API!! 1/n
Doc:
Github:
7
61
309
🌟gsplat==1.4.0🌟 () is now released!
- Supports 2DGS! (Kudos to
@WeijiaZeng1
)
- Supports Fish-eye camera! (Kudos to
@jefequien
)
- Prebuilt wheels for Windows! (Kudos to Martin de La Gorce)
2
26
232
🌟gsplat🌟 () now supports multi-GPU distributed training, which nearly linearly reduces the training time and memory footprint.
Now Gaussian Splatting is ready for the city-scale reconstruction!
kudos to this amazing paper:
3
24
190
#ECCV2024
Come and checkout our poster for 🌟 NeRF-XL🌟, which provides a principled solution for high fidelity city-level scale digital twins!
I'm not there due to visa issues but
@HavenFeng
will help present this work! Cheers for Haven!
Excited to release 🚀NeRF-XL🚀, a principled method to distribute your NeRF's parameters to multiple GPUs!
It is heuristics-free and mathematically equiv. with training your NeRF on a single (large) GPU.
Let's scale it up!🌎
17
81
324
1
14
133
3DGS-MCMC densification strategy is now officially supported with in 🌟gsplat🌟 with a one line code change!
kudos to
@hujh14
for bringing this into gsplat!
0
16
130
OMG. gsplat is on the 7-th of github trending today over ALL languages! And top1 in CUDA repos! Super glad to see people find it useful!
2
5
130
🌟gsplat 1.3.0🌟 now officially supports 3DGS compression. Easily get 15x compression rate in one line of code. Thanks for
@jefequien
to bring this in, and kudos to this paper:
A very compelling ranking on this leaderboard:
The
@nerfstudioteam
strikes again! This time gsplat is getting a massive compression boost, without sacrificing fidelity.
This translates to file sizes ~8mb per 500K Gaussians!
It's based on
@wimost
and
@FlorianIBarthel
's work, Self Organizing Gaussians.
🔗
3
13
93
2
12
124
Excited to share our ICCV23 paper!
The motivation behind this is better geometry modeling == better localization ability. And NeRF is perfect for extracting geometry from images.
Does NeRF only work for 3D reconstruction? 🙋♂️ NeRF-Det makes a novel use of NeRF for 3D detection, with large improvement while eliminating the heavy overhead of per-scene optimization.
website:
paper:
code:
0
12
67
2
7
72
It was fun to share to stories behind nerfstudio and gsplat!
Radiance Fields Episode 2 featuring Ruilong Li of Nerfstudio
0
11
43
3
6
52
The project page for our ECCV accepted paper:
Let’s make GPUs go brrr!
Want construct a ditigal twin at the scale of entire cities from a collection of images? Check out the NeRF-XL
@twominutepapers
video summarizing how!
Big props to co-authors
@ruilong_li
@akanazawa
@FidlerSanja
2
12
83
0
4
47
Thanks for having me! Excited to joining the party and share my thoughts about NeRF & 3DGS development!
Exciting news! Michael from
@RadianceFields
and I are thrilled to welcome
@ruilong_li
, one of the key driving forces behind gsplat development, as our next guest. Don't miss out - tune in next Thursday at 4 pm EST, 1 pm PT, and 10 pm CET!
2
8
115
0
1
42
Super cool! First time hearing about Slang.D but this auto differentiable property is super nice! Curious how the Slang.D compares with Triton / taichi etc which also supports auto diff with customized kernels.
I just released a Slang.D implementation of 3D Gaussian Splatting renderer.
This will allow researchers to write 3DGS papers even faster (sigh!) since there is no backward pass!
You can trivially change the forward pass like you would do on PyTorch!
6
48
282
1
5
43
And more would come!
0
0
37
Thanks
@RadianceFields
for writing this article. It has a nice interpretation on the use case of this technique. Existing era indeed!
NeRF-XL from
@ruilong_li
and the all star team of
@FidlerSanja
,
@akanazawa
, and
@frncswllms
brings multi GPU scaling from everyday captures to 25 kilometer cities.
🔗
1
12
100
2
3
32
Go gsplat!
🔥 Unleash the beast with gsplat v1.0! More features, blazing-fast training, and metrics so precise even INRIA would bow down. It splats so fast, Gauss himself would be jealous (and me, too)! Plus, it's Apache 2.0 licensed.
2
19
146
0
2
31
With this efficient backend, it's much easier to train a large scene with just a small GPU (8 GB requirement in this case).
*Source of the capture:
1
1
27
If you work on monocular reconstruction with NeRF, do check out our latest analysis paper! The analysis start from the data you captured, all the way to the sota methods and evaluation. Also practical tips are there to make things work better!
Do dynamic NeRFs work well on in-the-wild captures🤳? Not really there yet!
In NeurIPS 2022, we provide a reality check and show there's a large room to improve.
w/
@ruilong_li
,
@shubhtuls
, Russell,
@akanazawa
.
proj:
arXiv:
1/10
2
25
177
0
2
25
And it enables rendering a large scene with hundreds of millions of GSs in real-time.
2
0
21
Come and chat on Friday 10:30-12:30, at Foyer Sud 089!
Nerfacc w/
@ruilong_li
@hangg70
Matt Tancik
If you're into volumetric rendering perhaps you've heard of nerfacc! ⭐️Ruilong⭐️ can tell you all about how to unify all the existing sampling methods under the 🔎 of transmittance
Friday Morning!
1
0
8
0
2
20
And it is already been used by
@nerfstudioteam
to build the super cool NeRF all-in-one library nerfstudio:
1
0
19
This is a continuous dev. on top of the amazing gsplat v0.1.x from
@vickie_ye_
, from who I have learnt a lot! And a big thank to the guidance from
@akanazawa
and Matthew tancik, as well as the help from
@_maturk
, Jianbo Ye,
@justkerrding
,
@brenthyi
@pan_zhuoyang
(unordered)
0
1
19
Key word: In seconds!
Meet Genie, a research preview of an all-new kind of generative 3D foundation model
#MadeWithGenie
💬 Create 3D things in seconds on Discord
⚡️ Prototype in various styles
🎨 Customize materials
🆓 Free during research preview
Try it now 👇
84
269
2K
1
0
19
For those who have developed your Gaussian splatting application based on the official implementation, we also demonstrate via a fork of it, on how to seamlessly migrate to gsplat v1.0, with less than 100 lines code change, and get the best for free!
1
2
16
Gaussian rasterization is a sparse process, in which not all Gaussians are visible to the camera. So we leverage this sparsity and carefully engineer the rasterization as a "sparse operator", in which it only allocates minimal GPU memory that it needs.
1
1
16
Along with the CUDA backend that can be install simply via PyPI, we also provide a standalone script that reproduces the official metrics, but with much less training memory footprint and faster convergence.
1
0
14
Come and meet the team members
@nerfstudioteam
of the coolest NeRF toolbox ever at Siggraph Los Angeles!
🥳 We will be presenting nerfstudio at SIGGRAPH 2023 as a conference paper 🎉:
Myself and many of the
@nerfstudioteam
will be attending the conference! Looking forward to meeting the
@siggraph
community 🙂
4
18
158
0
1
15
My intern project
#AIChoreographer
from
@GoogleAI
is finally released! We created a huge 3D human dance motion dataset AIST++, and an algorithm to generate dance from music automatically. Thanks to collaborators
@shanyangmie
@akanazawa
and David Ross!
3
3
14
We also have a paper on arXiv to discuss the magic receipts. Check it out if you are curious:
1
1
13
Awesome project! Shout out for
@superbennyguo
who is a super good PhD at Tsinghua University.
Also I’m happy to see the nerfacc library I built a while ago can be helpful in this project!
1
0
12
Always love to see more engineering efforts in the field. AI won't evolve this fast without things like PyTorch/TF. And many problems actually can just be "simply" solved by some serious engineering efforts.
0
0
10
So realistic!
AI-driven 3D lifting has huge potential! Choose your favorite AI tool to generate/customize an avatar and bring it to life in 3D – something not possible with 3D scanning-based systems.
If you come to
@SIGGRAPH
, try a live demo as part of Emerging Technologies
#SIGGRAPH2023
2
100
397
0
2
8
Thanks to the team
@FidlerSanja
@akanazawa
@frncswllms
!
And the supports from
@NVIDIAAI
@berkeley_ai
0
1
8
Thanks for the article of introducing it! Happy weekend!
Ruilong and the
@nerfstudioteam
are starting the weekend with a bang!
It's the release of gsplat 1.0 and there are some exciting features!
🔗
🔗
0
5
42
0
0
8
Code entrance here:
0
0
7
“forget all the algorithms work, just take a fairly standard ConvNet, make it very big, train it on a big dataset (ImageNet), and just implement the whole thing in CUDA/C++. And it's in this way that deep learning as a field got a big spark.” - very well said
# CUDA/C++ origins of Deep Learning
Fun fact many people might have heard about the ImageNet / AlexNet moment of 2012, and the deep learning revolution it started.
What's maybe a bit less known is that the code backing this winning submission to the
165
900
7K
0
0
7
One of the papers I really like these days which deserves more attentions!
Excited to present our "Time Reversal Fusion (TRF)" this morning at
#ECCV2024
!!!
TRF enables an image-to-video model to generate towards a given end frame without any tuning 🚀
Come and chat with us at Poster 285 during the 10:30 - 12:30 session.
Video results and code
1
6
58
0
0
6
Super cool demo!!
11
127
778
1
1
5
@kwea123
哇感谢试用！这里想clarify一下nerfacc的design是user可以方便地pip install 并且集成到自己的代码里。nerfacc并不是一个codebase（需要user git clone 整个repo来使用）。自带的examples都是非常simple的code structure，仅仅为了展示如何使用这个repo。Again！thanks for all the feedbacks!!!
1
0
5
Special thanks to AIST Video Database created by
@shuhei2306
etc. With huge amount of AIST multi-view videos and our AIST++ 3D annotations, we hope to push forward the cutting-edge researchs about human analysis!!
0
1
4
@8Infinite8
@WeijiaZeng1
@jefequien
The paper is this: mesh exporting is not implemented unfortunately
1
1
4
Excited to release 🚀Nerfstudio Gradio WebUI🚀: Simplify your 3D model training & visualization with an intuitive GUI!
📊 Train models directly from your browser
👁️ Visualize with Viser
📦 Process data seamlessly
🚀 Export models with ease
And more...
2
9
31
0
1
4
@jiading_fang
@FidlerSanja
Thanks for bring this into our lidar! It looks pretty cool! We will cite it in our next revision!
0
0
4
This is how the game starts 🫡
Our Gemini model enjoyed watching yesterday's OpenAI event. Nice work,
@miramurati
,
@barret_zoph
and
@markchen90
(and everyone else behind the scenes)!
26
85
1K
0
0
3
Paper link:
Project page:
Dataset page:
1
0
2
@ErichLiang
awesome job!
HR-NeuS: Recovering High-Frequency Surface Geometry via Neural Implicit Surfaces
Erich Liang, Kenan Deng, Xi Zhang, Chun-Kai Wang
tl;dr: multi-resolution hash grid encoding+surface regularization+coarse-to-fine grid annealing
2
6
24
1
0
2
@psyth91
@akanazawa
@shuhei2306
Thank you Shunsuke!!! I’m really excited about this dataset as well!!
0
0
2
@taiyasaki
@jefequien
Yeah this benchmark we run is based on MCMC! Solid work! The way I see it is that any improvement towards better quality of 3DGS can also be analogous to compression, as one can achieve the same quality with less amount of Gaussians.
1
0
1
@ashwinkumar_v
Unfortunately this is not on our dev. roadmap, but you might find something useful from our community:
0
0
1
@linyoutian2
@kwea123
可以优化相机姿态的。rendering的gradient可以回传到ray进而回传到camera。ray marching的过程只是生成（t_start，t_end），这个过程不需要gradient并且也不会影响到相机优化的。
1
0
1
0
0
1
@oliver_batch
We follow the literature that down-res the outdoor scenes by 4x and indoor scenes by 2x. See script here
1
0
1
0
0
1
0
0
1
Amazing results! congrats!
@_TianyeLi
As the “hello world” to twitter, I'm excited to share with you "Neural 3D Video Synthesis"!
Arxiv:
Page (w/ full video):
@LvZhaoyang
, Mira,
@MZollhoefer
,
@simesgreen
,
@chlassner
,
@tanner_schmidt_
,
@_ChangilKim
, Steve and Michael
2
5
22
0
0
1
congrats!!!
@geopavlakos
CONGRATS
@geopavlakos
and your proud advisor
@KostasPenn
!!! It was such a pleasure collaborating with you in your formative research years!
@Penn
@KostasPenn
2
0
21
0
0
1