StableCascade_EmptyLatentImage_motorway_edition

Stable Cascade is a two-stage beast: a tiny, compressed stage C that does the composition, then a stage B that upscales that latent into the full-res image. If you've never built a Cascade workflow, the first shock is that there isn't one "empty latent" node - there are two latents to create. StableCascade_EmptyLatentImage makes both, blank, sized and compressed to your spec. This is the Motorway edition, so instead of two output wires you get both latents written into the pipe under keys.

The compression number is the weird part of Cascade and it's worth understanding. Stage C latents are heavily compressed - the default 42 means the effective resolution is roughly width/height divided by 42. That's the whole trick of the architecture: stage C thinks in postage-stamp coordinates and stage B inflates them back. You generally match whatever compression the checkpoint you're running was trained at, and the defaults (1024, 1024, 42) line up with the standard cascade checkpoints.

What to set

Because this node has no upstream connections, the MOTORWAY input is optional - leave it disconnected and it acts as a source/start node.

• width / height (default 1024 × 1024, step 8) - the stage C base resolution you're targeting.
• compression (default 42, 4–128) - the latent compression factor. Leave it alone unless you know your model wants a different one.
• batch_size (default 1) - how many blank latent pairs to create.
• OUTPUT_stage_c_key (default stage_c) - where the stage C latent goes in the Motorway.
• OUTPUT_stage_b_key (default stage_b) - where the stage B latent goes.

Output is the MOTORWAY with both latents stored. stage_c feeds the stage C sampler; stage_b feeds stage B's sampler.

Installing it

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

Restart; it's under agilly1989 Nodes → Motorway-ed → latent → stable_cascade. Manager has the pack. No pip dependencies - the wrapper is pure Python over ComfyUI's own node.

Where people get burned

Stable Cascade checkpoints are picky about resolution and compression. Going off the native resolution means stage B can produce tiles or artifacts, and changing compression away from what the checkpoint was trained with can silently wreck the image. Match the numbers your model card lists.

The Motorway half of the trap is the same as everywhere in this pack: keys must match exactly and are case-sensitive; 'MotorwayClass' object has no attribute 'hash_' means a key you referenced isn't in the pipe. And the pack itself is beta - the README says so in caps - with the auto-cloned motorway-ed nodes toggled on and off across releases. If an update hides the node, that's the pack, not you.