X-Portrait

Why it's ranked here

This is compelling research code for controlled portrait generation, with unusually visible preprocessing and conditioning mechanics. The practical burden is substantial: it requires a specific Python and CUDA setup, separate pretrained weights, and careful frame matching that can affect output quality.

What's good

The pipeline automatically searches for a driving frame whose landmarks best match the source pose. It separately preserves local evidence around both eyes and the mouth. LCM LoRA compatibility can reduce inference steps, while optional results from another reenactment method can strengthen expressiveness.

Tradeoffs

Setup requires Python 3.9, CUDA 11.8, many tightly constrained dependencies, and a separately downloaded checkpoint. Face detection failures trigger fallback behavior. Output quality can depend on precisely choosing the best matching frame, so weak inputs or poor alignment may require manual iteration.

How to use it well

Use it for research experiments where you control the source portrait, driving footage, GPU environment, and sampling settings. Start with clear faces and select a close pose match before tuning frame overlap or inference steps. It does not replace a hosted API, polished editor, or complete production service.

Observed

Primary language

Python

Runtime requirements

Python 3.9 and CUDA 11.8 are required by the README.

Installation surface

Shell-based environment installation with dependencies declared in requirements.txt.

Inference interface

The README documents shell-launched generation with source image, driving video, output, frame, overlap, and sampling parameters.

Model distribution

Pretrained weights must be downloaded separately and placed under the checkpoint directory.

License

No license is stated in the supplied repository text.