LTX2-Rapid-Merges

I'm winding down maintaining these models. I recommend checking out the merging script to create your own, as at least the NSFW model is very outdated. Thank you for all of the support and feedback!

These are experimental FP8 merges of the LTX2 diffusion model. "All in ones" do not seem to be in the cards, as LTX2 has lots of (large) moving parts, including 2 VAEs for video and audio. So, for the meantime, I'm just going to focus on the diffusion model. For all other models needed to run LTX2, please refer here: https://huggingface.co/Kijai/LTXV2_comfy

Workflow uploaded! It is messy as shit but does "everything" (I2V, First to Last Frame and T2V) very quickly and efficiently using all the tricks I've collected. I highly recommend using it and read the included note (particularly about required nodes). "Save output" is set to "false" on the VideoCombine, so switch that to "true" when you start making something you want to keep.

The text encoder can be found here (I recommend fp8-scaled): https://huggingface.co/Comfy-Org/ltx-2/tree/main/split_files/text_encoders

I can only recommend my workflow with the included settings, as there are so many pitfalls getting LTX to work properly. Old models have been removed because they were generated from a broken script. All models now should be "working" moving forward.

----- SFW VERSION HISTORY -----

V3: Distill LORA at 0.9 with the detailer LORA at 0.4. Good general purpose LTX2 model.

V4: Small adjustment using an updated merging tool. Includes the distill LORA at 1.0 and detailer at 0.5, which works well with my workflow.

V5: Now contains the I2V Adapter from https://huggingface.co/MachineDelusions/LTX-2_Image2Video_Adapter_LoRa. This seems to work just fine with T2V. I mixed only the video portion, leaving audio untouched (since the LORA degrades audio slightly otherwise).

----- NSFW VERSION HISTORY ----

V6: Contains every NSFW LORA I could find on CivitAI for LTX2 at varying weights at of 1/17/26. Fixed a major problem with these merges, and now this is actually a decent, legit first pass at a general purpose NSFW LTX2 diffusion model. Genitals overall are still weak, so I2V for best results. THIS DOES NOT INCLUDE MY LATEST POVNSFW

V41: The version number reflects all of the internal iterations before deciding that this was "good enough" for release! This combines my POV NSFW LORA with lots of great work by CJ, TenStrip and other creators on CivitAI. A new merging method is used that handles overlaped LORA tensors individually to bring the best of the LORAs together. USE WITH MY WORKFLOW

V62: Huge step up in quality from V41. Better order of NSFW LORAs. Includes the "Image2Video_Adapter_LoRa" for better I2V performance. Includes some new NSFW LORAs (but not my "pov v3 complete"). Significantly improves dialog and voice clarity (but not 100%). Use with my workflow

Please support LORA creators and special thanks to the authors of these!