We reimplemented with 2 collegues “SinGAN: Learning a Generative Model From a Single Natural Image” : the objective of SinGAN is to compare patch distributions across scales, enabling realistic texture and structure synthesis from a single input image. Stack: PyTorch, scikit-image, The technical aspect are that they used multi‑scale pyramids, adversarial training, we also study patch statistics.

Objective :

Reproduce and extend SinGAN with custom loss shaping for patch‑level fidelity.

Results :

Faithful single‑image generation and scale‑aware texture synthesis; ablations on patch loss variants.

Link to the github repository :

Here