EBGAN: A GAN-based Noise Model for Ultra-Low-Light EBAPS Image Simulation

High-gain sensors like Electron Bombarded Active Pixel Sensors (EBAPS) enable imaging in extreme low-light, but introduce complex noise that challenges existing synthesis models. In contrast to conventional CMOS sensors that linearly amplify an already noisy signal, EBAPS follows a sequential “Amplify-then-Add” cascade where dominant, physical signal amplification precedes the addition of electronic noise. However, current generative architectures model noise synthesis as a primarily additive process because they lack the structural priors needed to disentangle this physical cascade. To address this, we propose EBGAN, a generative network whose architecture is structurally constrained to mirror this sequential reality. We introduce a decoupled dual-pathway generator: a Conditional Amplified Gain Pathway (CAGP) explicitly models the non-linear signal amplification, while a Differentiable Physical Noise Decomposer (DPND) synthesizes the subsequent additive noise. Extensive experiments show our method achieves state-of-the-art performance on multiple fidelity metrics (MMD², PSD-Δ, NLF-RMSE). Crucially, its practical utility is validated by improving state-of-the-art denoisers by up to 2.5 dB in PSNR when our synthetic data is used for augmentation. Our work provides a robust, physically-grounded tool for generating realistic data, paving the way for improved image restoration on high-gain sensors.