Digital Camera Noise Synthesis

Playground

Sensor noise is not the white Gaussian speckle that denoising papers used to assume. It is coloured — brighter in the shadows than in the highlights, unequal across R, G and B, and correlated between neighbouring pixels by the camera’s own processing. Pick an image below and it is regrained as one of five smartphones at one of fourteen ISO levels, then drag the divider to compare against the clean original.

What runs here is not the network. The generator is conditioned on camera and ISO and trained adversarially; probing it as a black box reduces its behaviour to two measurable things: how loud the grain is at each clean intensity, and how strongly neighbouring pixels agree, and those are what this page ships: 70 presets of summary statistics, about 180 KB of numbers.

Every preset is reproduced exactly as measured — there is nothing here to dial in, so what you see is the model’s own answer for that camera at that ISO. Grain ×4 shows the noise field on its own, amplified four times about mid-grey; Seed draws a different realisation of the same distribution.

What the charts show

Both charts redraw for the selected preset, so they are the numbers behind whatever is on the canvas above — change camera or ISO and watch them move.

The model

n_c = k_c ∗ ( σ_c(I_c) · w_c )     noisy = clip(I + n, 0, 1) w is unit Gaussian noise, correlated across channels by a 3×3 Cholesky factor; σc is a 17-knot lookup evaluated at the clean intensity; kc is a 9×9 symmetric kernel supplying the spatial correlation, normalised so that Σk² = 1 — convolution then preserves exactly the variance the curve asked for.

Both ingredients are plotted above, and neither is what the textbook model would predict: the curve is not affine in intensity, and the kernel is not a blur. The paper is where they are derived and evaluated.

Citation

📄 A Generative Model for Digital Camera Noise Synthesis
arXiv:2303.09199 · PDF

Mingyang Song, Yang Zhang, Tunç O. Aydın, Elham Amin Mansour, and Christopher Schroers. 2023. A Generative Model for Digital Camera Noise Synthesis. arXiv:2303.09199 [cs.CV]. https://arxiv.org/abs/2303.09199

@article{song2023generativecameranoise,
  author  = {Song, Mingyang and Zhang, Yang and Ayd{\i}n, Tun\c{c} O. and Amin Mansour, Elham and Schroers, Christopher},
  title   = {A Generative Model for Digital Camera Noise Synthesis},
  journal = {arXiv preprint arXiv:2303.09199},
  year    = {2023},
  url     = {https://arxiv.org/abs/2303.09199}
}

Citing this page

The playground is not part of the paper — the parametric model, the 70 measured presets and this implementation were made for it. If those are what you used, please cite the page alongside the paper rather than instead of it.

@misc{song2026grainplayground,
  author       = {Song, Mingyang},
  title        = {Digital Camera Noise Synthesis: An Interactive Camera-Grain Playground},
  year         = {2026},
  howpublished = {\url{https://miya9756.github.io/mingyang-song.github.io/projects/grain/}},
  note         = {Accessed: 2026-07-28}
}

Plain BibTeX has no entry type for a web page, so one goes in as @misc with the address in howpublished; \url{} needs \usepackage{hyperref} or \usepackage{url} in your preamble. On biblatex you can use the real @online type instead, with separate url and urldate fields. Either way, set the access date to the day you actually visited.

Datasets

The noise model behind the presets was trained on an open-source smartphone-noise dataset, and the sample photograph comes from a standard test suite. Please cite the original authors: