EmptySD3LatentImage_motorway_edition

SD 1.5 and SDXL latents have four channels. SD3's latent has sixteen. That single number is why SD3-family models - SD3 Medium, and by extension the flow-matching architecture that Flux and friends run on - can't just reuse the classic EmptyLatentImage. The blank latent has to be built in the right channel layout or the model refuses to cooperate. EmptySD3LatentImage is that node, and this is its Motorway-ed version from agilly1989/ComfyUI_agilly1989_motorway.

The wrapper is pure plumbing: the core node builds the same zeroed 16-channel latent, then parks it on the Motorway pipe under a key instead of returning it as a socket.

The inputs

• width (default 1024, min 16, step 16) and height (default 1024, min 16, step 16) - pixel dimensions. The 16-step matters: SD3's VAE downscales 16×, so sizes that aren't multiples of 16 misalign. 1024×1024 is the square default; 1280×720 and 896×1152 are the common aspect ratios.
• batch_size (default 1) - images per batch.
• OUTPUT_LATENT_key (default LATENT) - the key the blank latent is stored under.

Output is the MOTORWAY pipe only. The sampler downstream picks the latent up via its INPUT_latent_key.

When to reach for it

Any SD3/flow-matching image checkpoint. If your workflow loads an SD3-based model - or a fine-tune of one - the sampler expects a 16-channel latent, and this is the only node in the pack that supplies it. If you plug in a 4-channel EmptyLatentImage instead, the model either errors on the channel mismatch or silently produces garbage.

One broader note that this whole pack sits inside: SD3 is where the flow-matching story starts, and flow models are exactly why people say "4–8 steps is normal now" and why aggressive schedulers like Karras hurt instead of help. None of that changes what this node does - it's still just a blank canvas - but it's the context for why SD3's own defaults (and this node's 1024 square default) look the way they do.

Install

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

Restart ComfyUI, or install via ComfyUI Manager under "ComfyUI_agilly1989_motorway." No extra dependencies; the model files are the download that actually matters.

Pack-level caveat, repeated for anyone landing here first: v1.1.7's own pyproject.toml reads "HOTFIX - REMOVED MONKEYPATCH AND GENERATED NODES," and the generated clones are commented out of __init__.py. Core Motorway nodes work; the _motorway_edition clones may not appear in your menu until the author re-enables them. If this one's missing, core EmptySD3LatentImage is identical - ramp its output onto the bus yourself with a Motorway 1x0.