Mohammad Mahdi Derakhshani

Multimodal Generative Models | Large-Scale Pretraining

Welcome! I'm Mohammad, a researcher in multimodal generative models, working across the full stack from large-scale data curation to multi-node training.

Most recently, I was a Senior Member of Technical Staff at New Theory, working on pretraining geometry-conditioned models for a generalist robot. Before that, I was a postdoc at the University of Amsterdam's VIS Lab and a community researcher at Cohere Labs, where I co-led NeoBabel. I completed my Ph.D. at the VIS Lab in 2026 with Cees Snoek and Yuki Asano; my thesis is On Post-Training Artificial Intelligence Foundation Models.

Earlier, I interned at Microsoft Research (2023) on spatial composition in text-to-image diffusion models and at Samsung AI Center (2022) on Bayesian prompt learning. I did my master's at the University of Tehran on object detection and image compression.

Research Interests

  • Multimodal generative models: pretraining and post-training of transformers, next-token prediction, diffusion, and flow matching
  • Large-scale training: data curation, scaling laws, multi-node training (FSDP, DeepSpeed)
Mohammad Mahdi Derakhshani

News

Tutorials & Lectures

Lecture slide
Scaling Transformers 5D Parallelism

Lecture, Foundation Models Course, UvA,

3D parallelism diagram
A Code-study Walkthrough of 3D Parallelism

Tutorial, Foundation Models Course, UvA,

Selected Publications

LiFT FID against inference FLOPs compared with dense DiTs
LiFT: Loop Flow Transformers

Mohammad Mahdi Derakhshani, Pedro M. P. Curvo, Gertjan J. Burghouts, Jan-Willem van de Meent, Cees G. M. Snoek.

arXiv bibtex

A looped Diffusion Transformer that can run far beyond its training depth. LiFT-L/2 achieves an FID 3.34 points lower than dense DiT-XL/2 with about 60% fewer parameters.

GIST semantic protection examples
Semantically Steerable Image Tokenization

Danilo de Goede, Mohammad Mahdi Derakhshani, Yuki M. Asano, Cees G. M. Snoek.

arXiv bibtex

An adaptive image tokenizer steered by a requested semantic fidelity, so it can discard pixel detail while preserving what the image depicts.

PlaySuite pipeline
PlaySuite: A Large-Scale Benchmark for Interactive Visual Intelligence

Dheeraj Varghese, Anna Vettoruzzo, Walter Simoncini, Michelle Lorena Acevedo Callejas, Mohammad Mahdi Derakhshani, Kristof Meding, Joaquin Vanschoren, Cees G. M. Snoek.

arXiv bibtex

A benchmark for interactive visual intelligence across more than 5K open-source video games, revealing a perception-action gap in fourteen recent open models.

Purrception method
Purrception: Variational Flow Matching for Vector-Quantized Image Generation

Răzvan-Andrei Matişan, Vincent Tao Hu, Grigory Bartosh, Björn Ommer, Cees G. M. Snoek, Max Welling, Jan-Willem van de Meent, Mohammad Mahdi Derakhshani*, Floor Eijkelboom*.

*: equal contribution.

ICLR bibtex

A hybrid flow model that samples like a continuous flow yet learns discretely, training faster than continuous and discrete flow-matching baselines.

NeoBabel results
NeoBabel: A Multilingual Open Tower for Visual Generation

Mohammad Mahdi Derakhshani, Dheeraj Varghese, Marzieh Fadaee, Cees G. M. Snoek.

arXiv bibtex

A fully open image generation model for six languages. It beats BLIP3-o 8B on m-GenEval (0.75 vs. 0.64) while being 2–4× smaller.

TULIP method
TULIP: Token-length Upgraded CLIP

Ivona Najdenkoska*, Mohammad Mahdi Derakhshani*, Yuki M. Asano, Nanne van Noord, Marcel Worring, Cees G. M. Snoek.

*: equal contribution; random order.

ICLR bibtex

Upgrades CLIP-like models to any token length, using relative position encodings and distillation of the original text encoder.

Bayesian prompt learning method
Bayesian Prompt Learning for Image-Language Model Generalization

Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees G. M. Snoek, Georgios Tzimiropoulos, Brais Martinez.

ICCV bibtex

Frames prompt learning as variational inference, improving generalization to unseen prompts across 15 benchmarks.

Other Publications

*: equal contribution; random order for TWIST & SCOUT.