roop-unleashed-wip — an actively developed fork with 3D pose matching, advanced masking, and expression warping

roop-unleashed-wip: more control over the swap, not just a swap button

roop-unleashed-wip is my ongoing fork of C0untFloyd's roop-unleashed — same core face-swap pipeline, but built out with a lot more control over how the final result actually looks, plus a bunch of stability fixes.

What's actually new here:

• 🖌️ A masking editor that follows the face — zoom/pan canvas with side-by-side live preview, and a manual mask that tracks the face frame-by-frame on video instead of falling apart the moment something moves.
• 👄 Surgical mouth control — restore the original mouth entirely, or just blend the inner mouth (teeth/tongue) back in while keeping the swapped lips. Small detail, huge difference in how "off" a swap feels.
• 🎭 Expression warp — nine presets (Happy, Sad, Angry, Surprised, and more) that push the swapped face toward an expression using TPS warping on landmarks already being tracked. No reference photo needed.
• ✂️ A real Editor tab — resize, rotate, reframe, retime, crop — then send it straight into Face Swap without leaving the app.
• 🧪 Experimental 3D source pose matching — an early attempt at warping the source to better match the target's head angle. Still rough on tricky angles, included for feedback more than as a finished feature.
• 🌑 Carbon Dark UI, plus a long list of stability fixes under the hood (batch crashes, mask quality, settings that actually persist between sessions).

Full changelog and feature breakdown: github.com/Adutchguy/roop-unleashed-wip

Getting it running is the easy part. Install through Pinokio and it detects your hardware and grabs the right build on its own — NVIDIA (CUDA/TensorRT), AMD (DirectML on Windows, ROCm on Linux), Apple Silicon, or plain CPU. Everything lives in its own isolated environment, nothing leaks into the rest of your system.

One ask: this is for technical and academic use. If you're swapping a real person's face, get their consent and label it as a deepfake before it goes anywhere public. Full disclaimer's in the README.

Still very much a work in progress (it's in the name) — bug reports, feedback, and PRs are all welcome.