Embedding connector has been moved from text encoder to diffusion model, so to continue using single files you now have to update KJNodes and use my loaders:
image
for GGUF:
image
22th January 2026: Tiny VAE and live previews:
LTX2 Tiny VAE has been trained by madebyollin, and now has preliminary support in ComfyUI nightly version.
It can be used with normal VAE Loader and encode/decode nodes, but the quality is very low so it's only useful for preview purposes.
Also currently for live animated sampler preview it can be only used with my LTX2SamplingPreviewOverride -node in KJNodes, simply load the VAE and plug it in, this overrides any preview setting too.
Turns out the video VAE in the initial distilled checkpoints has been wrong one all this time, which (of course) was the one I initially extracted.
It has now been replaced with the correct one, which should provide much higher detail
requires using updated KJNodes or new enough core ComfyUI VAELoader to work correctly
Separated LTX2 checkpoint for alternative way to load the models in Comfy
image
LTX2 GGUFs that include the metadata also work with the setup, if your GGUF nodes are up to date.
Dev and distill embeddings_connector weights are extracted from their respective original models, they do give quite different results, but it's unclear if there's a "correct" one to use in any given situation.
Loras folder includes rank reduced loras that can be useful to reduce the memory requirements.
