For the first time, we show that the Llama 7B LLM can be trained on a single consumer-grade GPU (RTX 4090) with only 24GB memory. This represents more than 82.5% reduction in memory for storing optimizer states during training.

Training LLMs from scratch currently requires huge

DoRA explores the magnitude and direction and
surpasses LoRA quite significantly

This is done with an empirical finding that I can't wrap my head around
NVIDIA AI
arxiv.org/abs/2402.09353
Shih-Yang Sean Liu Chien-Yi Wang Hongxu (Danny) Yin Pavlo Molchanov Min-Hung (Steve) Chen

ConvSSM: State Space Models for long videos
🎉 We finally released the code and the pretrained models.
Code: github.com/NVlabs/ConvSSM
Paper: arxiv.org/abs/2310.19694
NVIDIA AI Jimmy Smith

Join @NVIDIA and Dell Technologies today at 2:00 p.m. ET to hear from Target and Canadian Tire as they discuss the top #generativeAI retail use cases that are driving revenue, improving customer experiences, and boosting employee productivity. nvda.ws/48rGHa3 #NRF2024

Our next #ICCV2025 meeting will be in Honolulu, Hawaii 🌴

Meta spends twice more on R&D than Amazon, Alphabet, and Microsoft, and 4 times more than Apple, when normalized by revenue.
The only other big tech that comes close is Nvidia.
That tells you a lot about who is building the technology of tomorrow.

gpt-fast now supports mixtral-8x7B, in addition to gpt/llama.
1000 lines of simple pytorch code blazing it out!
github.com/pytorch-labs/g…

🚀 Is the future of vision models Synthetic? Introducing SynCLR: our new pipeline leveraging LLMs & Text-to-image models to train vision models with only synthetic data!
🔥 Outperforming SOTAs like DinoV2 & CLIP on real images! SynCLR excels in fine-grained classification &

Pre-trained diffusion models can be repurposed to estimate environment illumination from a single image!

I worked at Intel on Larrabee applications in 2007. Then I went to NVIDIA to work on ML in 2008. So I was there at both places at that time and I can say:

NVIDIA's dominance didn't come from luck. It came from vision and execution. Which Intel lacked.

Introducing FastViT, fast, super-small, general purposes vision transformers running at ~ 1ms on mobile. Code and pre-trained models are live.
github.com/apple/ml-fastv…
#ICCV2023

Introducing “VOODOO 3D: VOlumetric pOrtrait Disentanglement fOr One-shot 3D head reenactment”.
We present a real-time 3D aware one-shot head reenactment method that can generate consistent views from any angle MBZUAI ETH Zürich VinAI Pinscreen URL shorturl.at/gnr17

#Gauss ian Splatting based #Gauss ian Avatar of Carl Friedrich #Gauss

Did I say Gaussians a bit too many times? There is more. Happy to present our latest work #GAvatar by
Ye Yuan and Xueting Li

Project page: nvlabs.github.io/GAvatar/

Excited to share our text to animatable avatar project. TL;DR: primitive-based Gaussian representation + Implicit Gaussian attribute fields + normal/mesh extraction from Gaussians.
w/ Ye Yuan Yangyi Huang Shalini De Mello Koki Nagano Jan Kautz @umariqb

COLMAP-Free 3D Gaussian Splatting

paper page: huggingface.co/papers/2312.07…

While neural rendering has led to impressive advances in scene reconstruction and novel view synthesis, it relies heavily on accurately pre-computed camera poses. To relax this constraint, multiple efforts

Introducing “LitNeRF”

syntec-research.github.io/LitNeRF/

With a sparse, mobile, and easy-to-build capture setup, LitNeRF performs novel view synthesis and relighting by decomposing scene radiance into interpretable components that are motivated by physically-based rendering.
Google AR & VR

We are organizing the first Rhobin workshop on Reconstruction of Human-Object Interactions at #CVPR2024 with an amazing lineup of speakers. Join us and submit a paper with an extended deadline on March 20.
rhobin-challenge.github.io