← All projects

CVPR 2025

GenDeg

A diffusion model creates degraded versions of clean images, giving restoration models more varied examples to learn from.

Co-author

Restoration with synthetic training data

Haze restoration using the same PromptIR architecture. GD denotes training with GenDS. Ground truth is shown at right.

Generalization to another scene

Another published comparison of PromptIR with and without GenDS training.

How it works

Controllable paired data

GenDeg conditions a diffusion generator on a clean image, text and degradation statistics. It generated more than 550,000 paired examples across haze, rain, snow, motion blur, low light and raindrops, with control over degradation type and intensity.

From synthesis to restoration

Structure correction and intensity filtering retain suitable synthetic pairs. Combining these pairs with existing datasets forms GenDS, with more than 750,000 samples, for training all-in-one restoration models. The comparisons test the effect of this training data using the same restoration architecture.

Degradation synthesis and restoration

Train the degradation generator, generate and filter synthetic pairs, then train the restoration model.

Results

Training the same restoration model with GenDS improves several out-of-distribution results. LPIPS and FID measure different aspects of visual quality; lower is better for both.

Test set LPIPS Existing LPIPS GenDS FID Existing FID GenDS
O-Haze 0.254 0.165 109.9 74.6
LOLv1 0.241 0.167 112.0 73.1
RSVD 0.143 0.121 21.7 14.1
GoPro 0.198 0.170 31.7 36.2

CVPR 2025, GenDeg, Table 2

Lower is better for both metrics. Motion-blur FID does not improve.

Benefits vary by degradation. On GoPro, motion-blur LPIPS improves while FID worsens. The evidence compares the same Swin architecture with different training data.

Presentation

GenDeg restoration results across all six degradation types, including original inputs, baseline methods, GenDS-trained variants and ground truth.