LTX 2.3 - Best Face Swap + New Models! (Patreon)

Video - https://youtu.be/V25GORD4h_U

Overview

We now have Best Face Swap (BFS) LoRAs & nodes available for LTX 2.3, plus a host of new models available! This includes a v3 from Kijai (which works even better with custom audio, thanks to more bf16 blocks), plus FP8 distilled and NVFP4 models from Lightricks.

https://huggingface.co/Alissonerdx/BFS-Best-Face-Swap-Video

https://huggingface.co/Lightricks/LTX-2.3-nvfp4

https://huggingface.co/Lightricks/LTX-2.3-fp8

I've updated my LTX2.3 workflow so that it now includes:

Best Face Swap (BFS) group & BFS audio options

Optional ollama system prompt for BFS

Also updated ollama film director instructions per LTX2.3 guidelines

Single workflow - 3 loaders available. Delete what you don't want, keep what you do!

Kijai v3 file now available - better quality!

Lightricks official NVFP4 & FP8 distilled models now available

Replaced custom nodes with natives nodes where possible now (no ComfyMath or other custom maths nodes required)

Also basic ComfyUI native nodes have replaced the Radiance nodes by default (power users can add these back easily - or any other nodes - for pro upscaling, LUTs, etc).

Stage2 custom scheduler, steps & denoise

Workflow Information

Download links for the models files are available in the workflow. There are lots available, so download the ones you want to try!

Rather than have multiple different workflow files, I've included all the loaders for whichever set of files you want to use.

GGUF:

Checkpoint:

Or Kijai's split files:

Simply connect whichever model loader set you prefer (like in the image above) to the set nodes in the Init group. Delete (or mute) whichever loader groups you don't want.

Fewer custom nodes are also required now, thanks to Comfy-native maths nodes :)

Install custom nodes using ComfyUI Manager.

To run a workflow in ComfyUI, click the "Run" button.

Changes settings to match your requirements, such as the video length & resolution:

Help is included throughout the workflow in the various notes:

For those who love to tweak (like me) Stage 2 (Upscale) now also has extra options available in the Other Settings group:

Enable the Alt Sigmas STG2 if you're looking to experiment with those. By default this setting is off. This will give you very different outputs! I'm still playing around to see what the best combos are, but you can alter the steps & denoise too. Even 3 steps seems OK, which can save a bit of time on longer duration videos.

The relevant notes have been updated, but if using NVFP4 model, then the distill LoRA works best at full strength (or more)! I went up to 1.2 and got reasonable enough results:

The strength of 0.5 seems fine with the GGUF, however, so your mileage may vary based on your choice of model.

For BFS, IMG1 counts as the face image:

You'll need to enable the BFS ControlNet & video group:

Plus, don't forget to enable the BFS LoRA too:

Example input video:

Example output:

You can also use the ollama prompt enhancer:

Connect both the image and the alternative system prompt for the best results.

Note that the LoRA author has also provided a video analysis node as well which you may wish to add.

I've found you don't HAVE to use their exact format, and you can just type your required result with pretty good results!

If you want, the audio from the BFS video (if it has any) can also be piped through - or you can replace with your own custom audio at the flick of a switch.

Your choice of audio can also be set in the output group - the audio it generated or the original from the BFS input video:

If using the audio from the original BFS input video, it is a good idea pipe that in to LTX as well, as this way the model will still "see" the original audio (which helps with lip sync). Then, use the original audio from the BFS input video for your final video generation - if you want it to be exactly the same as the original (no changes from processing).

Videos which don't have any audio can, of course, use the generated audio.

Enjoy! :)