Roman Bachmann — ML Research Scientist

Hi! I'm Roman Bachmann, a Machine Learning Research Scientist at Apple. My research is focused on building scalable any-to-any multimodal foundation models for world modelling and visual reasoning. My goal is to build adaptable world priors that enable quick understanding of the environment and allow for global, out-of-sight reasoning.

Previously I was an EPFL PhD student at VILAB, advised by Amir Zamir. I received my M.Sc. degree in Data Science at EPFL, where I also completed my B.Sc. in Computer Science. During my studies, I interned as a research scientist at Apple and RIKEN AIP.

2026

Weblica: Scalable and Reproducible Training Environments for Visual Web Agents →

Oğuzhan Fatih Kar, Roman Bachmann, Yuanzheng Gong, Anders Boesen Lindbo Larsen, Afshin Dehghan

arXiv 2026

arXiv
Weblica: Scalable and Reproducible Training Environments for Visual Web Agents

MODUS: Decoder-only Any-to-Any Modeling of Diverse Modalities →

Mingqiao Ye*, Zhaochong An*, Zhitong Gao, Xian Liu, François Fleuret, Chuan Li, Amir Zadeh, Serge Belongie, Afshin Dehghan, Jesse Allardice†, David Mizrahi†, Oğuzhan Fatih Kar†, Roman Bachmann†, Amir Zamir

ICML 2026

arXiv Website Code Dataset
MODUS: Decoder-only Any-to-Any Modeling of Diverse Modalities

(1D) Ordered Tokens Enable Efficient Test-Time Search →

Zhitong Gao, Parham Rezaei, Ali Cy, Mingqiao Ye, Nataša Jovanović, Jesse Allardice, Afshin Dehghan, Amir Zamir, Roman Bachmann†, Oğuzhan Fatih Kar†

ICML 2026

arXiv Website Code
(1D) Ordered Tokens Enable Efficient Test-Time Search

VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization →

Andrei Atanov*, Jesse Allardice*, Roman Bachmann, Oğuzhan Fatih Kar, R Devon Hjelm, David Griffiths, Peter Fu, Afshin Dehghan, Amir Zamir

ICML 2026

★ Spotlight

arXiv Website Code
VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization

2025

How Well Does GPT-4o Understand Vision? Evaluating Multimodal Foundation Models on Standard Computer Vision Tasks →

Rahul Ramachandran, Ali Garjani, Roman Bachmann, Andrei Atanov*, Oğuzhan Fatih Kar*, Amir Zamir*

ICLR 2026

arXiv Website Code
How Well Does GPT-4o Understand Vision? Evaluating Multimodal Foundation Models on Standard Computer Vision Tasks

FlexTok: Resampling Images into 1D Token Sequences of Flexible Length →

Roman Bachmann*, Jesse Allardice*, David Mizrahi*, Enrico Fini, Oğuzhan Fatih Kar, Elmira Amirloo, Alaaeldin El-Nouby, Amir Zamir, Afshin Dehghan

ICML 2025

arXiv Website Code
FlexTok: Resampling Images into 1D Token Sequences of Flexible Length

2024

4M-21: An Any-to-Any Vision Model for Tens of Tasks and Modalities →

Roman Bachmann*, Oguzhan Fatih Kar*, David Mizrahi*, Ali Garjani, Mingfei Gao, David Griffiths, Jiaming Hu, Afshin Dehghan, Amir Zamir

NeurIPS 2024

arXiv Website Code
4M-21: An Any-to-Any Vision Model for Tens of Tasks and Modalities

ViPer: Visual Personalization of Generative Models via Individual Preference Learning →

Sogand Salehi, Mahdi Shafiei, Teresa Yeo, Roman Bachmann, Amir Zamir

ECCV 2024

arXiv Website Code
ViPer: Visual Personalization of Generative Models via Individual Preference Learning

2023

4M: Massively Multimodal Masked Modeling →

David Mizrahi*, Roman Bachmann*, Oguzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, Amir Zamir

NeurIPS 2023

★ Spotlight

arXiv Website Code
4M: Massively Multimodal Masked Modeling

Modality-invariant Visual Odometry for Embodied Vision →

Marius Memmel, Roman Bachmann, Amir Zamir

2023 Conference on Computer Vision and Pattern Recognition

arXiv Website Code
Modality-invariant Visual Odometry for Embodied Vision

2022

MultiMAE: Multi-modal Multi-task Masked Autoencoders →

Roman Bachmann*, David Mizrahi*, Andrei Atanov, Amir Zamir

2022 European Conference on Computer Vision

arXiv Website Code
MultiMAE: Multi-modal Multi-task Masked Autoencoders

CLIPasso: Semantically-Aware Object Sketching →

Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, Ariel Shamir

SIGGRAPH 2022

★ Best Paper Award winner

arXiv Website Code
CLIPasso: Semantically-Aware Object Sketching

2021

Omnidata: A Scalable Pipeline for Making Multi-Task Mid-Level Vision Datasets from 3D Scans →

Ainaz Eftekhar*, Alexander Sax*, Roman Bachmann, Jitendra Malik, Amir Zamir

2021 International Conference on Computer Vision

arXiv Website
Omnidata: A Scalable Pipeline for Making Multi-Task Mid-Level Vision Datasets from 3D Scans

2020

Training Binary Neural Networks using the Bayesian Learning Rule →

Xiangming Meng, Roman Bachmann, and Mohammad Emtiyaz Khan

2020 International Conference on Machine Learning

arXiv Code
Training Binary Neural Networks using the Bayesian Learning Rule

2019

Motion Capture from Pan-Tilt Cameras with Unknown Orientation →

Roman Bachmann, Jörg Spörri, Pascal Fua, and Helge Rhodin

2019 International Conference on 3D Vision

★ Selected for oral presentation

arXiv Dataset
Motion Capture from Pan-Tilt Cameras with Unknown Orientation

Global Motion Estimation from Pan-Tilt Cameras →

Roman Bachmann, Helge Rhodin, and Pascal Fua

2019 Central European Seminar on Computer Graphics (non-peer-reviewed)

★ Voted best presentation and third best paper

PDF Dataset
Global Motion Estimation from Pan-Tilt Cameras

Automatic 3D motion capture in alpine skiing using deep learning and computer vision →

Roman Bachmann, Helge Rhodin, Jörg Spörri, and Pascal Fua

8th Int. Congress on Science and Skiing (non-peer-reviewed)

★ Won first place in the Young Investigator Award competition

PDF Dataset
Automatic 3D motion capture in alpine skiing using deep learning and computer vision