Ayush Tewari (@_atewari) / X

Ayush Tewari

126 posts

Ayush Tewari

@_atewari

Assistant Professor @Cambridge_Eng

Joined December 2018

Pinned
Ayush Tewari
@_atewari
Feb 14
Is pixel prediction the best way to build a world model? Check out VDAWorld, an alternative path to building interpretable, editable, and physically grounded world models. We use a VLM to build a simulation of the scene with the help of a computer vision toolbox.
00:00
9.1K
Ayush Tewari
@_atewari
Dec 1, 2024
I am looking for graduate students for my new lab at the University of Cambridge! Join me to understand and build models of visual perception.
88K
Ayush Tewari
@_atewari
Sep 11, 2024
Excited to announce that I will be joining the University of Cambridge @Cambridge_Eng as an assistant professor in spring 2025! I will be looking for students for the next year. Check out @elliottszwu's thread for details on how to apply, and get in touch!
Elliott / Shangzhe Wu
@elliottszwu
Sep 11, 2024
I will be joining @Cambridge_Eng as an Assistant Professor in spring 2025, together with @_atewari. Clearly have been missing the good old UK rain after a wonderful year in California. Looking forward to opening this new chapter with brilliant colleagues and students!
49K
Ayush Tewari
@_atewari
Jun 18, 2023
After a very long wait, my VISA application for Canada to attend #CVPR was refused with reasons like "not satisfied that you will leave Canada at the end of your stay". Unfortunate.
52K
Ayush Tewari
@_atewari
Aug 26, 2023
Check out “Diffusion with Forward Models”. We learn to sample realistic 3D scenes from a single input image. Our models are trained on videos and do not require 3D training data! …ffusion-with-forward-models.github.io (1/n)
00:00
12K
Ayush Tewari
@_atewari
Jun 21, 2022
Check out our #CVPR2022 work Disentangled3D where we learn 3D GANs of objects with disentangled appearance and geometry components: vcai.mpi-inf.mpg.de/projects/D3D/
00:00
Ayush Tewari
@_atewari
Dec 17, 2020
Check out our work "i3DMM: Deep Implicit 3D Morphable Model of Human Heads". We build a morphable model with independent control over the identity, expression, and hairstyle components. Project Page: gvv.mpi-inf.mpg.de/projects/i3DMM/ Video: youtube.com/watch?v=4pYzV3…
Ayush Tewari
@_atewari
Dec 1, 2024
Replying to @_atewari
Apply here: postgraduate.study.cam.ac.uk/courses/direct…
9.7K
Ayush Tewari
@_atewari
Aug 26, 2023
Replying to @_atewari
I have been interested in learning 3D generative models from 2D data for a long time. Over the years, I have built models of faces, general objects, and now, scenes. Training using 2D data would enable us to learn at scale and build generalizable models. Stay tuned! (6/n)
466
Ayush Tewari
@_atewari
Dec 24, 2021
Replying to @BolkartTimo @vincesitzmann and @_krishna_murthy
Thanks, Timo!
Ayush Tewari
@_atewari
Jun 21, 2022
Replying to @_atewari
The key idea is to represent any object instance as a non-rigidly deformed canonical 3D volume. Our method learns the canonical volume, as well as its deformations, jointly during training. We can also use the deformations to compute dense correspondences between rendered images.
Ayush Tewari
@_atewari
Aug 26, 2023
Replying to @_atewari
This is joint work with Tianwei Yin, @GCazenavette, @eigenstate, @fredodurand, Bill Freeman, Josh Tenenbaum @MITCoCoSci, and @vincesitzmann. I have learned so much from everyone! Checkout Vincent's thread as well: x.com/vincesitzmann/… (5/n)
Vincent Sitzmann
@vincesitzmann
Aug 25, 2023
Introducing “Diffusion with Forward Models”, 𝗮 𝗺𝗼𝗱𝗲𝗹 𝘁𝗵𝗮𝘁 𝗰𝗮𝗻 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲 𝗱𝗶𝘃𝗲𝗿𝘀𝗲, 𝗿𝗲𝗮𝗹 𝟯𝗗 𝘀𝗰𝗲𝗻𝗲𝘀 𝗳𝗿𝗼𝗺 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗶𝗺𝗮𝗴𝗲, 𝘁𝗿𝗮𝗶𝗻𝗲𝗱 𝘄𝗶𝘁𝗵 𝗶𝗺𝗮𝗴𝗲𝘀 𝘄/𝗼 𝗮𝗻𝘆 𝟯𝗗 𝗱𝗮𝘁𝗮! …ffusion-with-forward-models.github.io 1/n
00:00
802
Ayush Tewari
@_atewari
Aug 26, 2023
Replying to @_atewari
There are a lot of uncertainties in this task. Deterministic methods, such as pixelNeRF, struggle to reconstruct regions not observed in the input image. (2/n)
00:00
383
Ayush Tewari
@_atewari
Aug 26, 2023
Replying to @_atewari
We introduce a diffusion model that tightly integrates the image formation process with the denoising network, allowing us to generate 3D scenes while only supervising in image space. (3/n)
349