Null-Space Diffusion Distillation for Efficient Photorealistic Lensless Imaging

TL;DR

This paper tackles the challenge of achieving photorealistic lensless imaging without ground-truth paired supervision. It analyzes diffusion-prior guidance and shows that measurement-space DPS can struggle with consistency, while reconstruction-space guidance via range–null decomposition (DDNM$+$) yields more plausible results. The authors introduce NSDD, a training-free, offline distillation of a $DDNM^+$ teacher into a single-pass student conditioned on $(\boldsymbol{y}, \boldsymbol{A}^{\dagger}\boldsymbol{y})$, achieving near-teacher perceptual quality with substantially reduced latency. Experiments on Lensless FFHQ and PhlatCam demonstrate competitive perceptual quality and faster inference than multi-step diffusion methods, indicating a practical path toward fast, ground-truth-free, photorealistic lensless imaging.

Abstract

State-of-the-art photorealistic reconstructions for lensless cameras often rely on paired lensless-lensed supervision, which can bias models due to lens-lensless domain mismatch. To avoid this, ground-truth-free diffusion priors are attractive; however, generic formulations tuned for conventional inverse problems often break under the noisy, highly multiplexed, and ill-posed lensless deconvolution setting. We observe that methods which separate range-space enforcement from null-space diffusion-prior updates yield stable, realistic reconstructions. Building on this, we introduce Null-Space Diffusion Distillation (NSDD): a single-pass student that distills the null-space component of an iterative DDNM+ solver, conditioned on the lensless measurement and on a range-space anchor. NSDD preserves measurement consistency and achieves photorealistic results without paired supervision at a fraction of the runtime and memory. On Lensless-FFHQ and PhlatCam, NSDD is the second fastest, behind Wiener, and achieves near-teacher perceptual quality (second-best LPIPS, below DDNM+), outperforming DPS and classical convex baselines. These results suggest a practical path toward fast, ground-truth-free, photorealistic lensless imaging.

Figures (6)

• Figure 1: Left: Displaying lensless captures, the reference (i.e., Ref.) image displayed to be captured, and the reconstructions using a Wiener filter and our proposed NSDD technique. Right: Comparison of perceptual quality (i.e., LPIPS) versus inference time on the lensless FFHQ dataset Neto_etal_SelfNeuralLensless_2025_OptReview for various diffusion-prior-based (i.e., DPS, DDNM$+$) and traditional (i.e., Wiener, ADMM) algorithms. Ours (i.e., NSDD) achieves high reconstruction quality at a fraction of the time required by other methods.
• Figure 2: Depiction of the two types of diffusion guidance approaches analyzed on this paper. On the top, the likelihood-score guidance with an under-steered trajectory (small gradient updates) and an over-steered trajectory (large gradient updates) for the same sample, showcasing the challenging tradeoff between consistency and quality. On the bottom, the proximal posterior-mean guidance approach enforces a projection onto a consistent path that leads to plausible reconstructions on the image domain.
• Figure 3: Parameter sweep for DPS and DDNM$+$ algorithms for samples of the lensless FFHQ dataset.
• Figure 4: NSDD overview. Top: Precomputation of the DDNM$+$ targets $T(\mathbf y)$ using a fixed seed ($\textbf{x}_T$). Bottom: Distill the concatenate measurement $\mathbf y$ with the range anchor $\mathbf A^\dagger\mathbf y$ along channel dimension, reduce 6$\to$3 channels via a small UNet $g_\phi$, and predict a null-space residual with the pretrained diffusion UNet $g_\theta$. The output $\hat{\mathbf x}_s=\mathbf A^\dagger\mathbf y+\hat{\mathbf x}_{\text{null}}$ is a single-pass reconstruction.
• Figure 5: Qualitative reconstructions for Lensless FFHQ and PhlatCam datasets. Left to right: measurement, Wiener, ADMM, DPS, DDNM$+$ (teacher), NSDD (ours), reference image (Ref.).