Original title: Benchmarking RAW and RGB Restoration in Image Signal Processors
Zihao Lu, Radu Timofte, Marcos V. Conde · September 2026
Digital cameras use a tool called an image signal processor, or ISP, to turn raw sensor data into normal color pictures. When photos are damaged by blur or noise, engineers want to fix them. Researchers asked a simple question: should you clean up the photo before or after this processor? They tested both setups using different smartphone types, two learned processors, and various repair models. They found that location alone is not the whole story. Fixing the raw data works better than using standard repair models on the final color image. However, the best results came from color repair models that were trained to understand the camera's processing steps. Knowing how the camera works makes the biggest difference.
You have a sketch and want to add colors to make it vibrant. You can color it before it's scanned and printed, or after. If you color before scanning, the scanner's settings might alter the hues. If you color after printing, the colors are fixed but might not blend well. In this paper, the camera's processing is like scanning and printing, and restoration is like coloring. The research shows that coloring after printing, while knowing how the scanner changes colors, results in the most vivid final picture.
For smartphone users and photographers, this means clearer images from the devices they already own. The findings help software developers build better editing tools that work with camera hardware.
The abstract indicates this is a benchmark study covering specific smartphone device groups, learned ISPs, and degradation types. It does not introduce new restoration models but evaluates existing ones. Results may not generalize to all camera systems or image issues.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Benchmarking RAW and RGB Restoration in Image Signal Processors — Zihao Lu, Radu Timofte, Marcos V. Conde
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.