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: 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.

This observation echoes recent works that proposed using other priors than denoisers, e.g. DPR and ShARP from @YuyangHu_666 et al., but also SNORE from M. Renaud. Yet, we obtain a different Tweedie formula, suggesting composing the restoration prior with its assoc. degradation.

We can thus use this prior just like any denoising prior in an iterative algorithms, i.e. use a super-resolution (SR) model as a prior for deblurring, an inpainting model for SR, etc. Below, I show SRx4 results with various priors: LAMA (inpainting), deblurring, SRx2.

Instead, composing a restoration model with the degradation for which it has been trained does yield a fixed point and a tweedie-like prior!. In other words, the restoration model should not be considered independently from the degradation for which it was trained.

🧵 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:

RT @TachellaJulian: I'll be in Singapore this week for #ICLR2025, presenting "UNSURE: self-supervised learning with Unknown Noise level and….

The code and demo are live!.🔗 Paper: 🔗 Code: 🔗 Demo:

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:

Despite not being unrolled, our network shows strong zero-shot generalization — even to unseen operators and noise levels!

We trained this network on multiple imaging tasks: motion blur, inpainting, MRI, CT, Poisson-Gaussian denoising, super-resolution. and across modalities (1, 2, 3 channels). It learned all tasks jointly — no task-specific retraining needed.

We propose two key architectural updates to make a UNet multitask and generalize:.✅ Inject knowledge of meas. operator A into inner layers (like conditioning in diffusion models). ✅ Share weights across modalities (grayscale, color, complex), adapting only input/output heads.

⚠️ But trained unrolled networks lose their theoretical properties anyway. So we asked: can we revive the UNet — and make it generalize?.

The issue: plain UNets don’t generalize well to new forward operators A. They’re also rarely used in this space because unrolled networks are thought to offer better interpretability.

Most architectures trained for solving y = Ax + e are tailored to specific operators A. But what if we used one universal UNet backbone?.No PnP, no unrolling, no retraining — just good inductive bias.

🧵 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 👇.

RT @TachellaJulian: 🚢🚢 deepinv v0.3.0 is here, with many new features! 🚢 🚢. Our passionate team of contributors keeps shipping more excitin….

RT @tobias_liaudat: ⚡ Last days to apply for this exciting project to make significant contributions to space-based telescopes like Euclid….

RT @TachellaJulian: 🚀 Do you want to learn about self-supervised learning for inverse problems? . ▶️ Check out the 3-hour tutorial present….

RT @ukmlv: “Typical priors: ‘This image looks natural.’. FiRE🔥priors: ‘This image survives degradation + restoration.’. Our CVPR 2025 paper….