Matthieu Terris

@cloneofsimo maybe check openaccess.thecvf.com//content/CVPR2…

Why do we not have any VAE trained with principles from stylegan3 literature? Seems like so many video artifacts can be resolved with anti-aliasing high frequency operations

@YuyangHu_666 Thanks Yuyang!

Fantastic work by Matthieu! It's incredible to see how general restoration models (beyond denoisers) can act as such a strong prior for solving general inverse problems. Huge potential here!

As a summary, don't forget to degrade your images before applying a restoration model 😁 For more details, come chat with me in Nashville! Last but not least, this is joint work with Thomas Moreau and @ukmlv 🫡 Paper: arxiv.org/abs/2411.18970 Code: next week!

But a cool feature inherited from this fixed-point approach is the ability to combine restoration models! Our results suggest that combining SR, deblurring and denoising priors works best.

🧵 I'll be at CVPR next week presenting our FiRe work 🔥 TL;DR: We go beyond denoising models in PnP with more general restoration (e.g. deblurring) models! A starting point observation is that images are not fixed-points of restoration models:

I'll be in Singapore this week for #ICLR2025, presenting "UNSURE: self-supervised learning with Unknown Noise level and Stein's Unbiased Risk Estimate" Ping me if you are around too!

@yuanzhi_zhu @HuraultSamuel @TachellaJulian definitely, I'll ping you next week!

@MatthieuTerris @HuraultSamuel @TachellaJulian Congrats! btw you are in Palaiseau? Can have a coffee some time

🧵 What if a single UNet could solve all inverse problems? In our latest preprint with @HuraultSamuel, Maxime Song, and @TachellaJulian, we build a single multitask UNet for computational imaging — and show it generalizes surprisingly well 👇 arxiv.org/abs/2503.08915

@YuyangHu_666 @HuraultSamuel @TachellaJulian Thanks Yuyang! I think it should be the case, but a significant challenge I'd see for the experimental validation is task heterogeneity

@MatthieuTerris @HuraultSamuel @TachellaJulian Interesting work. Can scaling law extends to imaging priors like these?

@ukmlv @HuraultSamuel @TachellaJulian Thanks! Not yet, sadly — this work is mostly empirical. But I think there'd be much to do!

@MatthieuTerris @HuraultSamuel @TachellaJulian Thanks for sharing! Are there any theoretical insights on the design of the universal UNet architecture?

The code and demo are live! 🔗 Paper: arxiv.org/abs/2503.08915 🔗 Code: github.com/matthieutrs/ram 🔗 Demo: huggingface.co/spaces/deepinv…

And in challenging real-world imaging problems where there’s no ground truth and limited samples, we can still fine-tune… ✨ fully unsupervised! ✨ Using recent ideas from equivariant imaging + SURE, we adapt the model to a single noisy image, e.g. on this tough SPAD problem: