Purrception: Variational Flow Matching for Vector-Quantized Image Generation

ICLR 2026

Purrception generates high-resolution images in vector-quantized latent spaces, sampled as continuous transport learned through discrete supervision.

Răzvan Andrei Matișan, 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 as last authors.

International Conference on Learning Representations (ICLR) 2026

DuoDiff: Accelerating Diffusion Models with a Dual-Backbone Approach

NeurIPS 2024

DuoDiff introduces a dual-backbone diffusion framework that accelerates image generation by using a shallow network in early sampling steps and a deeper one later, achieving faster inference with high-quality results.

Daniel Gallo Fernández*, Răzvan Andrei Matișan*, Alejandro Monroy Muñoz*, Ana Maria Vasilcoiu*, Janusz Partyka, Tin Hadži Veljković, Metod Jazbec

* Alphabetical order. Equal contribution.

Neural Information Processing Systems (NeurIPS) 2024

Workshop on Adaptive Foundation Models

NARAIM: Native Aspect Ratio Autoregressive Image Models

NeurIPS 2024

NARAIM introduces autoregressive vision transformers trained on images in their native aspect ratios, improving downstream classification accuracy by preserving spatial context and reducing distortions compared to square-resized inputs.

Daniel Gallo Fernández*, Robert van der Klis*, Răzvan Andrei Matișan*, Janusz Partyka*, Efstratios Gavves, Samuele Papa, Phillip Lippe

* Alphabetical order. Equal contribution.

Neural Information Processing Systems (NeurIPS) 2024

Workshop on Self-Supervised Learning Theory and Practice

Reproducibility Study of 'ITI-GEN: Inclusive Text-to-Image Generation'

TMLR 2024

This reproducibility study of ITI-Gen confirms its ability to improve inclusiveness and image quality in text-to-image generation, while revealing challenges with multiple attributes, proxy features, and scalability. It also proposes an alternative approach based on Hard Prompt Search to address these issues.

Daniel Gallo Fernández*, Răzvan Andrei Matișan*, Alejandro Monroy Muñoz*, Janusz Partyka*

* Alphabetical order. Equal contribution.

Transactions on Machine Learning Research (TMLR) 2024