Gabriel Raya

📍 The Netherlands

Where noise becomes structure. My research studies the regimes in diffusion models where uncertainty is resolved and structure becomes recoverable. These regimes are not uniform along the denoising trajectory. They shape diversity, memorization, guidance, and where training effort is useful. I develop this view through symmetry breaking in diffusion dynamics, information-guided noise allocation, and time-dependent guidance, aiming to replace heuristic choices in training and control with measurable structure along the denoising path.

I am a final-year Ph.D. researcher in Machine Learning at Tilburg University and JADS (TU/e), and a visiting researcher at the Donders Institute for Brain, Cognition and Behaviour under the guidance of Dr. Luca Ambrogioni and Prof. Dr. Eric Postma. My work combines probabilistic generative modeling, statistical physics, and information theory. I previously worked as a Research Scientist Intern on Sony AI's Foundational Models team in Tokyo, where I worked on information-guided training for generative models.

Structure formation

How diffusion trajectories move from noise to data through phases, attractors, stability changes, and symmetry breaking.

Information-guided training

Where uncertainty is resolved along the corruption path, which noise levels carry learning signal, and how schedules should allocate effort.

Generalization dynamics

How generative behavior emerges from data geometry, associative memory, and the transition from memorization to structured synthesis.

Control and robustness

How guidance, negative conditioning, efficient sampling, and diffusion-based correction behave under distribution shift.

Selected research

Symmetry breaking dynamics in diffusion models

NeurIPS

Symmetry breaking in diffusion dynamics

Diffusion generation passes through distinct dynamical regimes. Early trajectories collapse toward a central fixed point, while later dynamics break symmetry and move toward data-manifold attractors.

arXiv

Noise scheduling as information-guided allocation

Noise schedules allocate training effort before the distribution of denoising difficulty is known. We estimate where uncertainty is resolved and adapt training toward that profile.

arXiv

From memorization to generalization

Generative behavior can be analyzed through associative memory. Models move from storing examples to organizing them into structured, generalizing dynamics.

ICLR

Dynamic negative guidance

Negative prompting can fail when the reverse process is non-stationary. Dynamic Negative Guidance modulates guidance across time and state to remove unwanted content more accurately.

Before my Ph.D., I obtained an M.Sc. in Computer Science at Radboud University, where I worked on unsupervised out-of-distribution detection for digital pathology with Twan van Laarhoven and Jasper Linmans. I was also a visiting researcher at the Computational Pathology Group, Radboud UMC, working on unsupervised and representation learning for histopathology with Witali Aswolinskiy and Francesco Ciompi.

selected publications

Representative papers behind this line of work; full list on the publications page.

arXiv

Noise Scheduling as Information-Guided Allocation in Diffusion Training

Raya, Gabriel, Nguyen, Bac, Batzolis, Georgios, Takida, Yuhta, Stancevic, Dejan, Murata, Naoki, Lai, Chieh-Hsin, Mitsufuji, Yuki, and Ambrogioni, Luca

arXiv preprint arXiv:2602.18647 2026

Abs arXiv PDF

We study diffusion training noise schedules as an information-allocation problem and propose InfoNoise, a data-adaptive method that allocates training effort using entropy-reduction profiles estimated during training.
NeurIPS

Spontaneous symmetry breaking in generative diffusion models

Raya, Gabriel, and Ambrogioni, Luca

Advances in Neural Information Processing Systems 2024

Abs arXiv NeurIPS PDF Blog Code Poster Slides

Generative diffusion models have recently emerged as a leading approach for generating high-dimensional data. In this paper, we show that the dynamics of these models exhibit a spontaneous symmetry breaking that divides the generative dynamics into two distinct phases:1) A linear steady-state dynamics around a central fixed-point and 2) an attractor dynamics directed towards the data manifold. These two "phases" are separated by the change in stability of the central fixed-point, with the resulting window of instability being responsible for the diversity of the generated samples. Using both theoretical and empirical evidence, we show that an accurate simulation of the early dynamics does not significantly contribute to the final generation, since early fluctuations are reverted to the central fixed point. To leverage this insight, we propose a Gaussian late initialization scheme, which significantly improves model performance, achieving up to 3x FID improvements on fast samplers, while also increasing sample diversity (e.g., racial composition of generated CelebA images). Our work offers a new way to understand the generative dynamics of diffusion models that has the potential to bring about higher performance and less biased fast-samplers.
MIA

Diffusion models for out-of-distribution detection in digital pathology

Linmans, Jasper, Raya, Gabriel, Laak, Jeroen, and Litjens, Geert

Medical Image Analysis 2024

Abs Paper

We applied diffusion models for unsupervised out-of-distribution (OOD) detection in digital pathology, significantly outperforming existing methods in the Camelyon16 challenge. These models effectively identify OOD data with high accuracy and alter anomalous input to appear more normal. Despite not reaching supervised learning’s performance, diffusion models show considerable potential for anomaly detection in medical imaging.
arXiv

Memorization to Generalization: Emergence of Diffusion Models from Associative Memory

Pham, Binh, Raya, Gabriel, Negri, Matteo, Zaki, Mohammed J., Ambrogioni, Luca, and Krotov, Dmitry

arXiv preprint arXiv:2505.21777 2025

Abs arXiv PDF

We connect diffusion models to associative memory and show how memorization gives way to generalization as training progresses, offering a theoretical perspective on the emergence of generative behavior.
ICLR

Dynamic Negative Guidance of Diffusion Models

Koulischer, Felix, Deleu, Johannes, Raya, Gabriel, Demeester, Thomas, and Ambrogioni, Luca

International Conference on Learning Representations 2025

Abs arXiv OpenReview PDF

We introduce dynamic negative guidance for diffusion models, enabling more flexible and effective steering of the generative process.

news

May 13, 2026	I was recognized as a Golden Reviewer for ICML 2026.
May 25, 2025	I completed a research internship with the Sony AI Music Foundational Model Team.
Nov 10, 2024	Our workshop paper Memorization to Generalization: The Emergence of Diffusion Models from Associative Memory has been accepted as a contributed talk at SciForDL NeurIPS 2024!.
Nov 10, 2024	Our workshop paper Dynamic Negative Guidance of Diffusion Models: Towards Immediate Content Removal has been accepted at SafeGenAI NeurIPS 2024!.
Jan 10, 2024	Our paper Diffusion models for out-of-distribution detection in digital pathology has been accepted at the journal of Medical Image Analysis 2024!.