Puppeteer Humans, Swap Their Clothes, Spin the Camera

What it does

Impersonator++ (iPERCore) implements a Liquid Warping GAN with attention for human image synthesis. Feed it a source image and a target pose, and it renders a new photo-realistic frame; it can also transplant a person’s clothes onto another body or rotate the virtual camera around them. The system targets 512×512 output and can reach 1024×1024 if your GPU has the headroom.

The interesting bit

Rather than siloing motion imitation, appearance transfer, and novel view synthesis into separate models, the authors unified all three under one attention-driven architecture. It is pitched as a more generalizable, higher-resolution successor to an earlier ICCV project, though the README is light on exactly what “more powerful ability in generalization” means in practice.

Key highlights

• Single pipeline handling motion imitation, appearance transfer, and novel view synthesis together.
• Supports bullet-time effects and multi-view outputs, turning a still setup into a virtual camera rig.
• Resolution scales from 256×256 up to 1024×1024, with a published VRAM budget table for each stage.
• Modernized for PyTorch 1.8+ and RTX 30-series GPUs.
• Includes a Google Colab notebook for cloud-based testing.

Caveats

• VRAM requirements escalate quickly: 9.8 GB for 512×512 personalization and 20 GB for 1024×1024, so high-res work is effectively restricted to workstation or data-center cards.
• The paper is listed as “currently under review” at IEEE TPAMI, meaning the project is academic work still awaiting final publication.

Verdict

Computer-vision researchers and hobbyists building human-synthesis demos should bookmark this; casual users without high-end NVIDIA hardware or patience for research code should probably skip it.