HunyuanVideo Empty Latent

Video latents aren't image latents. A HunyuanVideo clip is a stack of frames in a 32-channel latent space, and the frame count has to follow the model's temporal rules - you can't just ask for any number. HunyuanVideoEmptyLatent builds that empty video latent for you: width, height, frame count, batch, and it produces the correctly-shaped tensor the HunyuanVideo sampler starts denoising from. It's the video-generation equivalent of an "empty latent image" node, sized for this specific model.

Boring, essential, wire-it-once. But getting the shape right is what stands between you and a cryptic tensor error, so it's worth understanding what it's enforcing.

How it works

It allocates a zeroed 32-channel latent for the resolution and length you specify. The frame count is the interesting constraint: it defaults to 21 and steps by 4, which reflects how the video VAE compresses time - frames come in a 4n+1 cadence (1, 5, 9, … 21, …), so the node steps in fours to keep you on valid values. That's the same temporal-compression math the whole Wan/Hunyuan family shares; if you've ever wondered why your clip came back a couple of frames short of what you asked, this rounding is why the good nodes make you pick a legal length up front.

Defaults sit at 848×480 - 480p is HunyuanVideo 1.5's comfortable, fast resolution, especially with the step-distilled model - and you scale up from there as your hardware allows.

The inputs and outputs that matter

• width / height (default 848×480, step 16) - clip resolution. 480p is the sweet spot for speed; higher costs a lot more VRAM and time.
• frames (default 21, step 4, max 257) - clip length in frames. The step-4 cadence keeps you on the model's valid 4n+1 frame counts. More frames means a longer clip and more compute.
• batch_size (default 1) - how many clips at once.

Output: latent - straight into your HunyuanVideo KSampler.

How to install it

ComfyUI Manager: search ComfyUI-QwenImageWanBridge, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-QwenImageWanBridge

then restart. No model download for the node itself - it builds a tensor. The HunyuanVideo 1.5 diffusion model, VAE, and Qwen2.5-VL encoder live elsewhere in the graph.

Common issues & troubleshooting

Your clip is a couple of frames shorter than requested. That's temporal VAE compression, and it's why the node steps frames by 4. Stick to the values the field offers (21, 25, … 257) rather than fighting it; those are the lengths that decode cleanly.

Channel-mismatch error at the sampler. You probably used a generic empty-latent node (wrong channel count) instead of this one. HunyuanVideo needs the 32-channel latent this node produces.

It's slow / OOM at higher resolution. Video is genuinely GPU-tier-dependent - VRAM and time scale with both resolution and frame count. Generate at 480p with a modest frame count first; the step-distilled HunyuanVideo 1.5 model at 8-12 steps is what keeps 480p fast on a single 4090. Upscale a good clip afterward rather than generating big up front.