EmptyHunyuanImageLatent

EmptyHunyuanImageLatent is the starting blank for Hunyuan Image - Tencent's image model, not to be confused with Hunyuan Video. Same org, same license umbrella, completely different latent: this node builds [batch_size, 64, height // 32, width // 32], a 64-channel tensor downsampled 32× per side. That's the most aggressive compression in this whole family - 32× means a 2048×2048 image starts as a 64×64 latent, smaller than a thumbnail, carrying 64 channels of encoded meaning.

The inputs are the standard trio: width (default 2048, step 32, range 64–16384), height (same), and batch_size (default 1, up to 4096). The 2048 default and the 32-step are both telling you the model's training sweet spot - Hunyuan Image 3.0 was trained at high resolutions, and the node's step of 32 matches the latent grid (2048 ÷ 32 = 64). Keep your canvas in multiples of 32 or you'll be fighting the grid.

The honest context

Hunyuan Image 3.0 is the model behind this node: an 80B image MoE with 13B active per token, which the community measured at roughly 320GB of VRAM with quantization off the table. That's the fact that defines everything about this node's audience - it's a real, capable model with a permissive-enough Tencent license, and almost nobody can run it. It's not a bug in the node; it's the hardware reality. You'll reach for this if you have a serious card or a cloud box and want Hunyuan's strengths (strong prompt adherence, big resolutions) without touching the video line. If you're on consumer VRAM, this node is a curiosity - the "Hunyuan" in the name will keep pulling you in, and the model card will send you back out.

What it isn't

It's easy to mix up with the video side of the house. EmptyHunyuanLatentVideo and EmptyHunyuanVideo15Latent serve Hunyuan Video (and the 1.5 rebuild); this node serves the image model, and the two don't share a latent space - one is a flat 2D 64-channel tensor, the other is 5D video. If a Hunyuan Video workflow tells you to use a blank and this node shows up in search, check which model your checkpoint is before wiring it.

Also note the contrast with its sibling EmptyHunyuanLatentVideo-family defaults: Hunyuan Image assumes a big square canvas, and there's no aspect-ratio machinery here - you type the resolution, the node makes the blank, the sampler does the rest.

It ships with ComfyUI core - no install. The checkpoint is the download that matters, and the VRAM is the tax. If you can run it, it's a solid blank; if you can't, now you know why everyone talks around it.