BFS Best-Face-Swap-Video video model

Alissonerdx/BFS-Best-Face-Swap-Video is a community LoRA adapter built on top of Lightricks’ LTX-2.3 family. It has real commercial use, but the license still limits it. 

 LTX-2.3 is an open weights DiT audio video model from Lightricks, a company known for Facetune and other AI editing tools. Lightricks presents it as a local friendly video model with text to video, image to video, video to video, audio to video and synced audio generation.

The developer, Alissonerdx also known as Alisson Pereira Anjos, already has related work on Hugging Face. That includes BFS-Best-Face-Swap for images and other LTX and LoRA tests.

This seems like one of the early signs of how far LTX-2.3’s IC-LoRA setup can be pushed outside the main control uses. Instead of using IC-LoRA mostly for depth, pose or motion tracks, this project tries to use first frame identity conditioning to keep a swapped head steady through the clip. The idea is smart-ish and the early reaction from the community looks upbeat. The release matters because it pushes LTX closer to a video identity editing workflow, not just a video model.

It still looks like a test stage project. The repo warns about identity leakage and failures around cuts, and the docs say prompt text is not doing much yet. So this does not look like a ready production tool. It looks more like solid proof that LTX-2.3’s control setup can handle a focused video face swap pipeline.

Early talk from the community looks positive, with Reddit and YouTube posts reacting to the release. The project got quick attention on Hugging Face and fits with LTX-2.3’s newly documented IC-LoRA setup. That makes it worth watching... but not fully mature or fully proven yet.

The BFS project makes a video head swap result through LTX-2.3 workflows. The repo says to do a high quality head swap on frame 0, use that frame as the conditioning input, then run the video generation.

For the base model, LTX says 720p and 1080p are available in hosted versions on the LTX site with Fast and Pro API options. Higher resolutions can run locally up to 4K, but they need a lot more VRAM. IC-LoRA guidance points to 704×1216 at 24–30 FPS as a good range for control driven work.

The repo also lists known limits. Identity leakage. Hard cuts can reset identity. Detailed prompting has little effect right now. And the trigger phrase is just head swap.