Discrete PSF Priors Improve Blind Aberration Correction

Blind lens aberration correction is hard because the degradation is spatially varying, lens-specific, and usually unknown at inference time. Earlier LensLib pre-training pipelines train a universal restoration model on synthetic lens libraries, but the paper argues that two bottlenecks limit transfer: biased lens-source distributions and restoration networks that lack an explicit optical prior. FoundCAC addresses both with a larger sampled lens library and a discrete latent representation of point spread functions.

Figure 1 frames the change: the authors move from a one-pass blind restoration model trained on a biased lens library to a prior-guided pipeline trained on AODLibpro, with latent PSF features retrieved inside the restoration network.

Core Contribution

The central idea is to preserve the blind setting while still giving the model a physically grounded degradation signal. FoundCAC does not require the target lens PSF at inference. Instead, it learns a Latent PSF Representation (LPR): a vector-quantized codebook built from synthetic PSF maps during pre-training, then predicted from degraded images during blind restoration. This turns PSF use from a direct measurement requirement into a retrieval problem over a learned discrete degradation dictionary.

The other contribution is AODLibpro, a lens library built to reduce sampling bias. Rather than sampling mainly from lens design parameters or RMS spot radius, the authors classify lenses by optical image quality across fields of view. They divide degradation severity into strong, medium, and mild classes, and spatial variation into six classes based on per-FoV OIQ trends. The resulting train/test split uses non-overlapping lenses and a uniform sampling basis across severity and spatial patterns.

Technical Approach

FoundCAC is trained in three stages. Stage I uses degraded images, clear targets, and ground-truth PSF maps. A PSF encoder maps PSFs into continuous latent features, then vector quantization assigns each feature to a nearest codebook element. The restoration network receives degraded-image features fused with quantized PSF features and is optimized with reconstruction and codebook-alignment losses.

Stage II removes the PSF input. The learned codebook is frozen, and a blind PSF predictor estimates latent PSF features directly from the degraded image. Those features are quantized by the same codebook before being fused into the image restoration backbone. The paper’s ablation suggests this discretization is the useful part: unconstrained continuous PSF-feature prediction underperforms because it entangles degradation cues with image content.

Stage III adapts FoundCAC to a target lens. The codebook remains frozen while the predictor and restoration network are fine-tuned. The authors evaluate both full-shot adaptation and few-shot adaptation using only 5% of target data and 10K fine-tuning iterations, which is the practical case if collecting paired target-lens data is expensive.

Results and Analysis

On RealLens-Sim, FoundCAC reports the best average score among the compared blind pipelines: 26.84 dB PSNR, 0.852 SSIM, and 0.1585 LPIPS. The closest LensLib-PT baseline listed, AODLibpro-PT with SwinIR, reaches 25.91 dB, 0.843 SSIM, and 0.1793 LPIPS; the fast two-step method is lower at 24.38 dB, 0.742 SSIM, and 0.2786 LPIPS. The gain is not uniform across all optical regimes, but it is largest where the paper’s premise matters most: severe minimalist optical systems and misalignment cases.

The LPR mechanism is also supported by model-paradigm ablations. A plain blind baseline scores 28.46 dB PSNR on AODLibpro Test; direct PSF prediction reaches 28.25 dB, and direct continuous PSF-feature prediction reaches 28.42 dB. The GT-guided discrete VQ representation improves this to 28.97 dB, while the full blind FoundCAC pipeline reaches 29.12 dB. That pattern supports the authors’ interpretation: the codebook is not only a compression device, but a constraint that makes the blind inverse problem better posed.

The adaptation results are more practically interesting than the zero-shot numbers. Under few-shot adaptation, FoundCAC pretraining improves Single-Lens-I from 27.22 dB for scratch training without PSF to 30.90 dB, and improves the smartphone lens case from 32.07 dB to 33.34 dB. In the depth-aware setting, it reaches 26.25 dB with 5% of target data, close to its 26.75 dB full-shot result and above scratch training with GT PSF supervision at 23.69 dB. This is credible evidence that the frozen LPR codebook transfers information useful for lens-specific fine-tuning.

Caveats

The main caveat is evaluation realism. Real-captured images lack paired full-reference ground truth, so the strongest numerical claims come from simulated or real-design simulation settings. The paper adds no-reference RealLens-Snap metrics and visual comparisons, but those metrics can penalize conservative restorations or reward image-content changes. The authors also state that the learned codebook comes from synthetic refractive LensLib data, which may limit coverage for fabricated lenses, assembly tolerances, sensor/ISP effects, diffractive optics, depth-coupled degradations, under-display interference, low-light conditions, and sensor noise. FoundCAC is best read as a well-tested framework for PSF-prior-guided blind correction, not as a solved general camera pipeline.