iiinova — where your digital self keeps evolving

lebellig @lebellig.bsky.social · 23/09/2026

New interest: super-resolved images from undertrained flow matching models

061

lebellig @lebellig.bsky.social · 23/09/2026

I should keep this checkpoint for later

110

lebellig @lebellig.bsky.social · 02/09/2026

We’ve released the InSARFlow checkpoints, trained on two types of data: 🌎 Real SAR interferograms from Mexico City 🧪 Synthetic dataset + we added a simple notebook to make it easy to get started and experiment with InSARFlow 🧑‍💻 Joint work with Arnaud Breloy and @nshaud.bsky.social

000

lebellig @lebellig.bsky.social · 02/09/2026

We’re excited to release the updated article and code of InSARFlow! 🌊⛵ InSARFlow is a Riemannian flow matching model designed to denoise SAR interferograms while preserving the cyclical nature of phase differences.🌀 📄 Article: hal.science/hal-05710871... 💻 Code: github.com/lebellig/ins...

1113

lebellig @lebellig.bsky.social · 19/08/2026

Yesterday I watched a movie about an AI4Science researcher who gets angry after having his research grant application rejected (badly explained movie plot). Maybe we should take it as a warning about research funding cuts

241

lebellig @lebellig.bsky.social · 24/07/2026

NeurIPS submissions confirmed to be a heat-loving species. Warmer year, bigger bloom. Every degree we add, the deadline gets denser 🌻 Good news for the field, we're having a really good growing season 👨‍🌾

092

lebellig @lebellig.bsky.social · 21/07/2026

We show that increasing the alignment score at inference time improves source fidelity, leading to generated images that better preserve the source content, even when the model is trained on image pairs with low/medium/high alignment (see figure).

100

lebellig @lebellig.bsky.social · 21/07/2026

The key idea is to condition the bridge on image-pair alignment, providing a simple and effective way to improve source fidelity at inference time. We explore both metadata-based alignment and embedding-based alignment scores, enabling our method to be applied even when metadata are unavailable

100

lebellig @lebellig.bsky.social · 21/07/2026

The last project of my PhD is finally out! 🪴 It was a pleasure collaborating with Aimi on this work! We introduce A²BM: Alignment-Aware Bridge Matching, a new framework for image-to-image translation with weakly aligned image pairs. Paper 📄: arxiv.org/pdf/2607.16294

1144

lebellig @lebellig.bsky.social · 03/07/2026

Heading to #ICML2026 🇰🇷 and interested in diffusion models, flow matching, and their generalization capabilities? Don't miss the excellent tutorial by @mathurinmassias.bsky.social and @quentinbertrand.bsky.social on Monday! 📍 Hall D1 🗓️ Monday, July 6 🕘 9:00–11:30 AM Details: icml.cc/virtual/2026...

2116

lebellig @lebellig.bsky.social · 28/04/2026

My PhD thesis manuscript will be available in the coming months, but I’ve written two blog posts based on the related work chapter: 1. Generative modeling with flow-based models 🪚 2. Data-translation with flow and diffusion bridges 🔨 Open to feedback and discussions! lebellig.github.io/blog/

1357

lebellig @lebellig.bsky.social · 28/04/2026

Delighted to have successfully defended my PhD thesis on "Generative models for Earth Observation, from denoising to domain adaptation" 🪴 This wouldn’t have been possible without the support of my colleagues, my PhD advisor, and my family and friends. Thank you all!

4275

lebellig @lebellig.bsky.social · 18/03/2026

Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis by Hila Chefer et al. (arxiv.org/abs/2603.06507). New SSL loss for flow matching that encourages meaningful representation learning without relying on an external visual encoder for alignment. Improves generation on many modalities.

030

lebellig @lebellig.bsky.social · 11/03/2026

The Spacetime of diffusion models: an information geometry perspective by Rafał Karczewski et al. (arxiv.org/abs/2505.17517) blog: rafalkarczewski.github.io/blog/2026/di... Geodesics in the (xt, t) spacetime of diffusion models -> new distance between clean data points + transition path sampling!!

060

lebellig @lebellig.bsky.social · 22/01/2026

Meta Flow Maps enable scalable reward alignment, Peter Potaptchik et al. (arxiv.org/abs/2601.14430) This article introduces Meta Flow Maps: a stochastic generalization of consistency models (one-step generation) that allows efficient reward steering at inference time or during fine-tuning.

050

lebellig @lebellig.bsky.social · 08/12/2025

I use drum mic kits for punchier presentations

010

lebellig @lebellig.bsky.social · 18/11/2025

Calling it for today... I tried using the gemini 3 Pro preview to build some js animations, and it went well

030

lebellig @lebellig.bsky.social · 18/11/2025

Sorry but it's getting pretty when leaving python

000

lebellig @lebellig.bsky.social · 18/11/2025

Interpolation between two gaussian distributions on a flat torus (my personal benchmark for new llms)

150

lebellig @lebellig.bsky.social · 18/11/2025

"Back to Basics: Let Denoising Generative Models Denoise" by Tianhong Li & Kaiming He arxiv.org/abs/2511.13720 Diffusion models in pixel-space, without VAE, with clean image prediction = nice generation results. Not a new framework but a nice exploration of the design space of the diffusion models.

0132

lebellig @lebellig.bsky.social · 12/11/2025

"Curly Flow Matching for Learning Non-gradient Field Dynamics" @kpetrovvic.bsky.social et al. arxiv.org/pdf/2510.26645 Solving the Schrödinger bridge pb with a non-zero drift ref. process: learn curved interpolants, apply minibatch OT with the induced metric, learn the mixture of diffusion bridges.

041

lebellig @lebellig.bsky.social · 31/10/2025

Great article! But can the preference score go up to 2? You know, because 1 just isn’t aesthetic enough.

030

lebellig @lebellig.bsky.social · 30/10/2025

“Entropic (Gromov) Wasserstein Flow Matching with GENOT” by D. Klein et al. arxiv.org/abs/2310.09254 Transport between two distributions defined on different spaces by training a noise-to-data flow models in the target space, conditioned on the source data and leveraging Gromov–Wasserstein couplings

031

lebellig @lebellig.bsky.social · 14/10/2025

Diffusion Transformers with Representation Autoencoders by Boyang Zheng, et al (arxiv.org/abs/2510.116...) Unexpected result: swapping the SD-VAE for a pretrained visual encoder improves FID, challenging the idea that encoders' information compression is not suited for generative modeling!

0151

lebellig @lebellig.bsky.social · 10/10/2025

"How to build a consistency model: Learning flow maps via self-distillation" by @nmboffi.bsky.social et al (arxiv.org/abs/2505.18825) New method to train flow maps without any pretrained flow matching/diffusion models!

040

lebellig @lebellig.bsky.social · 08/10/2025

"Be Tangential to Manifold: Discovering Riemannian Metric for Diffusion Models" Shinnosuke Saito et al. arxiv.org/abs/2510.05509 High-density regions might not be the most interesting areas to visit. Thus, they define a new Riemannian metric for diffusion models relying on the Jacobian of the score

030

lebellig @lebellig.bsky.social · 10/09/2025

Grateful for the opportunity to speak at tomorrow’s Learning Machines seminar (RISE+@climateainordics.com) on generative domain adaptation and geospatial foundation models benchmarking for robust Earth observation 🌍 Join on Sept 11 at 15:00 CET! www.ri.se/en/learningm...

041

lebellig @lebellig.bsky.social · 30/08/2025

Late to the party but I like the fact that you can use geodesic random walk (like really simulating the random walks) to derive the SDEs necessary for diffusion models on Riemannian manifolds (from arxiv.org/abs/2202.02763)

070

lebellig @lebellig.bsky.social · 21/08/2025

I'll be at #GRETSI in Strasbourg next week! Friday morning, I'll present our work on Riemannian flow matching for SAR interferometry (generation and denoising) 🛰️ Also really looking forward to the poster sessions and all the exciting conferences on the program! 📄 hal.science/hal-05140421

150

lebellig @lebellig.bsky.social · 18/08/2025

Anyone aware of a cats --> pure evil creatures image translation benchmark? not even neural networks’ dreams reached this level of nightmare fuel

140

lebellig @lebellig.bsky.social · 07/07/2025

New episode in this line of work from @giannisdaras.bsky.social et al. on training diffusion models with mostly bad/low-quality/corrupted data (+few high-quality samples). This time for proteins! 📄 Ambient diffusion Omni: arxiv.org/pdf/2506.10038 📄 Ambient Proteins: www.biorxiv.org/content/10.1...

132

lebellig @lebellig.bsky.social · 01/07/2025

Added to my reading list: Adjoint Schrödinger Bridge Sampler by Guan-Horng Liu et al. arxiv.org/abs/2506.22565

040

lebellig @lebellig.bsky.social · 24/06/2025

Drop the "conditional", just "flow matching", it's cleaner.

2101

lebellig @lebellig.bsky.social · 11/06/2025

I was intrigued by "Mean Flows for One-Step Generative Modeling" and, in particular, how it handles averaging the marginal velocity field during training. In practice, they don't and replace it with the conditional one in their loss function. I wonder how mismatches impact generation...

Link to the article: https://arxiv.org/pdf/2505.13447v1

030

lebellig @lebellig.bsky.social · 21/05/2025

"Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling" by Michal Balcerak et al. arxiv.org/abs/2504.10612 I'm not sure EBM will beat flow-matching/diffusion models, but this article is very refreshing.

072

lebellig @lebellig.bsky.social · 29/04/2025

"Probability Density Geodesics in Image Diffusion Latent Space" by Qingtao Yu et al. arxiv.org/abs/2504.06675 They propose an algorithm to traverse high-density regions when interpolating between two points in a diffusion model latent space.

020

lebellig @lebellig.bsky.social · 28/04/2025

TerraMind (previous post), the new geospatial generative model from IBM and ESA, was trained on the TerraMesh dataset. Great to see large-scale datasets with aligned modalities! Now waiting for the release :) 📄 arxiv.org/abs/2504.11172

010

lebellig @lebellig.bsky.social · 28/04/2025

IBM and @esa.int introduce TerraMind, a new geospatial generative model. With 'Thinking in Modalities', they generate missing modalities during model finetuning. I'm curious about its modality translation capabilities 👀 📄 arxiv.org/abs/2504.11171 🐍 huggingface.co/ibm-esa-geospatial

041

lebellig @lebellig.bsky.social · 26/04/2025

I really liked this approach by @matthieuterris.bsky.social et al.They propose learning a unique lightweight model for multiple inverse problems by conditioning it with the forward operator A. Thanks to self-supervised fine-tuning, it can tackle unseen inverse pb. 📰 arxiv.org/abs/2503.08915

071

lebellig @lebellig.bsky.social · 25/04/2025

Very cool article from Panagiotis Theodoropoulos et al: arxiv.org/abs/2410.14055 Feedback Schrödinger Bridge Matching introduces a new method to improve transfer between two data distributions using only a small number of paired samples!

042

lebellig @lebellig.bsky.social · 19/03/2025

It reminds me of this result (from arxiv.org/abs/2104.11222), you can finetune the jpeg compression level of your training set to have the best FID on church images.

141

lebellig @lebellig.bsky.social · 13/03/2025

"Inductive Moment Matching" by Linqi Zhou et al. I like the use of multiple particles to apply a loss similar to consistency models, but on distributions. Training is stable and gives high-quality generated images in very few sampling steps 📄 arxiv.org/abs/2503.07565 🌍 lumalabs.ai/news/inducti...

072

lebellig @lebellig.bsky.social · 25/02/2025

Nice research work from @nicolabourbaki.bsky.social et al. Enhances latent generative models by regularizing the VAE's latent space with an equivariance loss. The finetuning process is straightforward + demonstrates improvements in just 5 epochs! 📄 arxiv.org/abs/2502.09509 🐍 github.com/zelaki/eqvae

091

lebellig @lebellig.bsky.social · 11/02/2025

I've been using this time sampling trick to stabilize diffusion model training for many years, yet I don't recall reading a comparison with naive uniform sampling + I rarely see it in flow matching codebases. Do you use it?

131

lebellig @lebellig.bsky.social · 07/02/2025

the only definition of straightness I'm aware of for such applications comes from the rectified flow article (arxiv.org/pdf/2209.03003) but there might be others

010

lebellig @lebellig.bsky.social · 07/02/2025

would say that Schrödinger Bridges trajectories are straight but not smooth (img from arxiv.org/pdf/2409.09347 App E.1). diffusion SDE: not straight and not smooth diffusion ODE/flow matching: not straight but smooth flow matching with reflow: straight and smooth SB: straight but not smooth

340

lebellig @lebellig.bsky.social · 05/02/2025

Now that bluesky has gifs (it didn't work?), I can share (again) my educational notebook on discrete flow matching (by Itai Gat et al.). Also please check the original article and official implementation by Meta! 🐍 github.com/gle-bellier/... 🐍 github.com/facebookrese... 📄 arxiv.org/abs/2407.15595

1152

lebellig @lebellig.bsky.social · 13/01/2025

What was the first paper you read when you started doing research? 🕵️ For me, "DDSP: Differentiable Digital Signal Processing" (by @jesseengel.bsky.social et al.) when I was an intern at IRCAM working on expressive synthesis of musical performances. (2020... I'm a rookie) 📄 arxiv.org/abs/2001.04643

061

lebellig @lebellig.bsky.social · 06/01/2025

At this time last year, we were working on using diffusion models to detect out-of-distribution satellite images 🛰️ We tested this in several contexts, including flood image detection! 📄 Article: arxiv.org/abs/2404.12667 🌍 CVPR EarthVision poster: gle-bellier.github.io/pdfs/posterE...

0141

lebellig @lebellig.bsky.social · 05/01/2025

"The Superposition of Diffusion Models using the Îto Density Estimator" (@martaowesyou.bsky.social, @lazaratan.bsky.social et al.) It's nice to see an easy-to-compute log-likelihood estimator for SDE sampling of diffusion models (not just ODE) 📄 arxiv.org/abs/2412.17762 🐍 github.com/necludov/sup...

0224