Mingqiao Ye

PhD Student in Computer Science

EPFL

I am a PhD student in Computer Science at EPFL, VILAB, advised by Prof. Amir Zamir. My research scales one model across many modalities and tasks, rather than building a separate specialized model for each.

MODUS, my recent work, is a single decoder that generates any modality from any other (RGB, depth, segmentation, text, and learned features alike), with no modality-specific heads, losses, or pipelines. What interests me is what a single unified model makes possible: modalities that strengthen each other when learned together, and the same model fluent across perception, generation, and reasoning. Earlier, I built widely-used systems for high-quality and open-world visual understanding, including HQ-SAM, Gaussian Grouping, Cascade-DETR, and EntitySAM.

I received my M.Sc. from ETH Zurich, where I worked with Dr. Lei Ke and Dr. Martin Danelljan, and my B.Sc. from Zhejiang University. I previously interned at Adobe Research with Dr. Joon-Young Lee.

Publications

* denotes equal contribution

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

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

ICML 2026

A single decoder that treats diverse modalities symmetrically, from RGB, depth, and segmentation to text and deep features, with no modality-specific heads, losses, or task pipelines. Two experts (autoregressive for discrete tokens, flow matching for continuous latents) enable any-to-any generation in one shared causal context.

[PDF] [Project Page]