Test-Time Optimization of Physical ISP Parameters from a Single Calibration Pair for Cross-Camera Adaptation

Every camera carries an image signal processor (ISP) that converts raw sensor data (RAW) into a viewable sRGB image; each body needs its own tuning. An ISP trained on one body degrades on another, and adapting a learned ISP measures how far it falls short, not why. We adapt a hybrid ISP instead: a deep-learning controller predicts the seventy named physical coefficients of a differentiable model-based pipeline, so what the pipeline can express is known before measurement. One calibration pair (RAW, sRGB) refits those coefficients to an unseen camera; the ISP then deploys alone, with no added weights and no per-image network pass. The one color transform the pipeline cannot express is predicted before measurement, and so is where its error appears. The five unseen bodies are four DSLRs and one smartphone. Refitting all coefficients to each test image’s own ground truth sets the best the pipeline can reach, and one pair recovers most of it. The predicted deficit appears on one body in the predicted band; the measurement decides which body. A closed-form color stage fitted from the same pair recovers +3.28 dB of peak signal-to-noise ratio (PSNR) there. A gate that reads only calibration data applies it to that body alone. The deployed ISP runs at 341× fewer multiply–accumulates per pixel than a weight-space reference retrained by us. The corrected fit still trails that reference; the remainder is unattributed.