These are experimental FP8 merged versions of the LTX2 diffusion model. An "all-in-one" merge seems unlikely because LTX2 contains many (large) components, including two VAEs for video and audio. So, for now, the focus is solely on the diffusion model. For all other models required to run LTX2, please refer here: https://huggingface.co/Kijai/LTXV2_comfy

The workflow has been uploaded! Although it's a bit messy, it performs "all" operations (I2V, first-to-last frame conversion, and T2V) very quickly and efficiently. Please read the accompanying instructions (especially the section regarding required nodes).

VideoCombine's "Save Output" is set to "false" by default, so when you start generating content you want to keep, please switch it to "true".

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

It is highly recommended to use the provided default settings for the workflow, as there are many pitfalls to getting LTX running properly.

Old models have been removed as they were generated by a broken script.

Now all models should work properly. All NSFW LoRA models suitable for LTX2, with varying weights, updated on January 17, 2026.

Fixed a major issue during the merging process, and now this is indeed a good, legitimate first draft of a general NSFW LTX2 diffusion model.