ModelMergeLTXV_motorway_edition

LTX-Video's diffusion backbone is a transformer with its own block layout - caption projection, a single-stream adapter, and a run of transformer_blocks - and ModelMergeLTXV is ComfyUI's per-block merge node for it. This is its Motorway-ed clone. Video-model merging is younger and rarer than the SD1/SDXL scene, but the mechanism is the same weight-patch interpolation, and the node gives you the same per-region control.

The block layout

• patchify_proj - the input projection that turns the latent patch into transformer tokens.
• adaln_single and caption_projection - the adaptive norm and text-caption conditioning paths. These carry a lot of the prompt-following behavior.
• transformer_blocks.0–transformer_blocks.27 - the 28 main transformer blocks, the bulk of the model. Lower blocks are earlier in the network (structure and motion semantics); higher blocks are where fine temporal detail and texture land.
• scale_shift_table and proj_out - the output projection back to latent space.

Every ratio defaults to 1.0, and the family gotcha applies: untouched, the merge returns model2. Set a base blend across all the floats first, then push regions - raise caption_projection if you want one parent's prompt adherence, nudge the higher transformer_blocks for temporal detail.

Why you'd merge LTX at all: LTX's selling point is speed - the community's line is "it might often not be as good as WAN, but you can do ten generations in the time" - and the model family has grown (0.9.x, 13B, distilled variants) with noticeably different strengths. Merging lets you combine a quality checkpoint's look with a distilled variant's speed without retraining. One real convenience: this node only touches the diffusion_model, so you never have to worry about the VAE/CLIP side of the merge - they pass through untouched.

The inputs

• INPUT_model1_key (default model1) and INPUT_model2_key (default model2) - the two LTX-Video models.
• The block ratios, each 0.0–1.0, step 0.01, default 1.0.
• OUTPUT_MODEL_key (default MODEL) - where the merged model lands.

The wrapper keeps every ratio as a widget, turns the models into key strings, and keyed-writes the result. Runtime behavior matches stock ModelMergeLTXV.

Install

One install for the whole pack. ComfyUI Manager, search "ComfyUI_agilly1989_motorway", or:

cd ComfyUI/custom_nodes
git clone https://github.com/agilly1989/ComfyUI_agilly1989_motorway

Restart ComfyUI. No requirements.txt, no model downloads.

Troubleshooting

• Result is pure model2? All ratios default to 1.0 - set a base blend first.
• Merging a distilled LTX variant with a non-distilled one is the risky direction: distilled models expect different sampling (fewer steps, no CFG), so the merged weights may inherit conflicting assumptions. Sample with low steps and see.
• Merge LTX with LTX only - block counts differ across video models (Mochi has 48 blocks, Cosmos has its own layout).
• 'MotorwayClass' object has no attribute 'hash_' = a model key isn't in the Motorway; check the loaders ran and the key spelling matches (case-sensitive).
• Pack caveats: BETA, and the newest published version (1.1.7) disabled clone generation, so on a fresh install these _motorway_edition nodes may be missing from the menu. They're present in the 1.1.0–1.1.6 builds comfy.icu indexed.