Empty HunyuanVideo 1.5 Latent

This node is the blank latent for Tencent's HunyuanVideo 1.5 - the November 2025 follow-up that cut the model to 8.3B parameters and dropped the local VRAM floor dramatically versus the original 13B release. The shape it allocates is a good tour of what changed between the two versions, because the latent got deeper and the spatial compression got more aggressive.

What it creates

[batch_size, 32, t, height/16, width/16]. Compare that to the original HunyuanVideo latent, which is 16 channels at height/8 and width/8. 1.5 doubled the channel count and doubled the spatial compression. The trade is worth understanding: more channels means more of the video's information survives into the latent, and the 16x downscale means the tensor is physically smaller per pixel - so the model can chew bigger videos without exploding VRAM, at the cost of pushing more work onto the decoder.

Temporal compression is unchanged: t = ((length - 1) // 4) + 1, so length steps by 4 and the defaults sit at 848x480x25 - the same friendly video size the rest of the family uses. The node also tags the latent with downscale_ratio_spacial: 16 so downstream nodes know how to map latent space back to pixels.

Where it sits

Straight into the latent input of a HunyuanVideo 1.5 workflow: prompt → this latent → sampler → VAE decode. You'll see it paired with the 1.5 image-to-video nodes and the newer 1.5 super-resolution and latent-upscale machinery, where the latent gets upscaled between passes rather than round-tripped through pixels.

The input you'll actually touch

Width, height, length, batch_size - the same four-knob layout as every video latent node, with steps (16, 16, 4) doing the shape-math for you. The one trap: don't feed a 1.5 latent into a v1 sampler or vice versa. The channel count is hard-coded into the tensor shape, and a 32-channel latent handed to a model expecting 16 produces immediate garbage, not a subtle error. Keep the empty latent matched to the checkpoint family you loaded.

Worth knowing

HunyuanVideo's community standing is complicated - it got to the local-video party first in late 2024, then Wan showed up and took the ecosystem, and 1.5 arrived too late to win the crowd back. But it's still actively developed by Tencent and the 1.5 line runs on far more modest hardware than the 45-60GB the original wanted, which is exactly why this node and its siblings exist in core. Ships with ComfyUI; nothing to install, just pick the right HunyuanVideo checkpoint to go with it.