ModelMergeLTXV

LTX Video is the speed tier of local video generation - Lightricks' family that drafts clips in seconds where Wan takes minutes. The community's relationship with it is the draft-in-LTX, refine-in-Wan workflow, not "let me blend two LTX fine-tunes together." So this node, honestly, is an edge case of an edge case: merging video models is rare, and merging LTX models is rarer. But it's here, it ships with core, and if you've got two LTX checkpoints you love it's the tool.

How it works

The family engine again: clone model1, mix model2's weights in per-block with ratio as the fraction of model2, longest-prefix matching on the weight keys. LTX's transformer is a DiT with adaLN conditioning, which is why the sliders read the way they do:

• patchify_proj. - the patch embedding that turns video tokens into the transformer's input space
• adaln_single. - the adapted layer-norm modulation, where timestep conditioning gets injected
• caption_projection. - the text-conditioning projection
• transformer_blocks.0. – transformer_blocks.27. - the 28 attention blocks themselves
• scale_shift_table and proj_out. - the output side

And the usual trap: every slider defaults to 1.0, meaning "take model2." Run without touching anything and you get model2 wholesale, not a blend. Put your base in model1, the flavor in model2, and lower the sliders you actually want to mix.

The gotchas that matter for video

Both inputs must be the same LTX variant with the same block layout - this node was written for the original 2B LTX-Video architecture (it landed November 2024, before the 13B and LTX-2). Feed it mismatched variants and the weight keys won't line up, the merge matches nothing, and you silently get model1 back. Same for the "must be the same architecture" rule that runs through the whole merge family.

The output is a MODEL that feeds the usual video sampler pipeline - VAE encode, sample, VAE decode - and it behaves exactly like either input would, because only weights changed.

Should you?

If you're on this page, you've probably already got the two LTX models and a reason. Fine. Start with small merges (0.3–0.5 on the transformer blocks), test on a short clip, and expect the video-model version of the universal merge truth: blending similar-aesthetic models is consistent, blending opposites is a coin flip. And if you were hoping this would let you merge LTX with Wan - it won't. Different architectures entirely. Merging across video families isn't a thing.