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

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.