EmptyHunyuanLatentVideo_motorway_edition

HunyuanVideo was the December 2024 moment when local video generation felt real - Tencent's 13B model, uncensored, surprisingly cinematic, and demanding an absurd 45–60GB of VRAM. This node makes the empty latent tensor that model denoises into a clip. It's the Hunyuan-specific sibling of the other Empty*LatentVideo nodes in this pack, tuned to HunyuanVideo's own compression: 16 latent channels, an 8× spatial downscale, and a 4× temporal downscale.

This is the Motorway-ed version from agilly1989/ComfyUI_agilly1989_motorway, so instead of a LATENT output socket it writes the blank latent onto the pipe under a key you name.

The inputs

• width (default 848, step 16) and height (default 480, step 16) - frame size in pixels. The defaults are Hunyuan's comfortable 16:9 starting point; 1280×720 and 1920×1080 are the usual bumps, and VRAM decides whether they're realistic.
• length (default 25, step 4) - number of frames. Because the temporal axis compresses 4×, (length - 1) // 4 + 1 frames are actually stored. 25 frames at ~8fps is about 3 seconds, which is the kind of clip HunyuanVideo defaults to.
• batch_size (default 1) - clips per batch. Leave it at 1.
• OUTPUT_LATENT_key (default LATENT) - the key the blank latent is stored under on the Motorway.

The only output is the MOTORWAY pipe. The latent itself travels by key: a motorway-ed sampler later reads it through its INPUT_latent_key.

Where it goes

Text-to-video with a HunyuanVideo checkpoint: store the latent early, then let the sampler denoise it. If you're on the 1.5 generation instead (8.3B, ~14GB floor), note that it uses a different latent layout (32 channels, 16× spatial), and this node targets the original 1.0 layout - another reason to check which checkpoint your workflow actually loads. HunyuanVideo 1.5 is the one people run locally in 2026; the 13B original mostly survives in archives and old workflows.

Install

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

Restart ComfyUI, or grab it via ComfyUI Manager under "ComfyUI_agilly1989_motorway." No extra Python deps; the real download is the HunyuanVideo model + VAE in models/checkpoints and models/vae.

The pack-level caveat applies here too: v1.1.7's own pyproject.toml says "HOTFIX - REMOVED MONKEYPATCH AND GENERATED NODES" and __init__.py has the clone registration commented out, so this _motorway_edition node may not appear in your menu even though the core Motorway nodes work. Beta pack, honest sign on the door, fix in progress.