Turn a photo into a puppet, a mannequin, or a new camera angle

What it does impersonator is the reference PyTorch implementation of the ICCV 2019 Liquid Warping GAN paper. Feed it a source image of a person plus a reference image or sequence, and it synthesizes new frames in which the source mimics the reference’s pose, wears the reference’s clothes, or is seen from a new viewpoint. The project is geared toward researchers working with the iPER and DeepFashion datasets, though it accepts arbitrary internet photos.

The interesting bit Instead of training three separate specialists, the project treats motion imitation, appearance transfer, and novel-view synthesis as a single unified framework. The authors also issued a public correction to their original LPIPS metric interpretation—an unusually direct admission that their first arXiv table used an inverted similarity score.

Key highlights

• Three built-in demo modes: motion imitation, appearance swap, and novel view synthesis.
• Runs on relatively modest hardware, requiring a minimum of 3.8 GB GPU memory.
• Supports custom inputs beyond the bundled iPER and DeepFashion samples.
• Includes evaluation metrics and training documentation for the iPER dataset.
• Explicitly tested only on Linux; other platforms are unsupported.

Caveats

• Platform support is narrow: the code is tested only on Linux, and the README explicitly warns that other platforms are untested.
• External dependencies are brittle: pretrained weights, checkpoints, and sample data must be downloaded manually from OneDrive or BaiduPan links.
• The dependency stack—PyTorch 1.2, torchvision 0.4, and CUDA 10.0—reflects its 2019 vintage and may need compatibility tweaks on modern environments.

Verdict Worth a look if you are reproducing 2019 human-synthesis benchmarks or building pose-driven video pipelines. Skip it if you need a maintained, cross-platform library with modern PyTorch support—this is research code with dated dependencies.