Empty HunyuanVideo 1.0 Latent

HunyuanVideo made local video generation feel real back in December 2024 - cinematic, uncensored, trainable - and every HunyuanVideo 1.0 workflow starts with this exact node. Empty HunyuanVideo 1.0 Latent creates the empty video latent that the 13B model denoises into your clip. It's the video-model equivalent of Empty Latent Image: a zeroed tensor with the right channel count, spatial compression, and temporal compression baked in, so the sampler knows what it's working with.

Four inputs:

• width / height - default 848×480, step 16. That's the classic launch-era resolution (480p); the model's other common bucket is 1280×720 at 60GB+ VRAM territory. Step must be a multiple of 16.
• length - default 25, the number of frames, step 4. That's a short clip - roughly one to three seconds depending on your frame rate; most community workflows run 121 or 201 frames. The well-known 201-frame setting gives you a long loop and, per the community, a free looping artifact.
• batch_size - default 1, how many clips at once.

Output: a LATENT shaped [batch, 16, (length-1)/4 + 1, height/8, width/8] with downscale_ratio_spacial: 8. Those numbers are the model's contract: 16 latent channels, 8× spatial compression, 4× temporal compression. Every 4 frames of video become one latent frame - which is why length steps by 4.

Where it fits

Standard T2V graph: this node → KSampler with Hunyuan conditioning → VAEDecode → Save WEBM. For I2V you'll instead use the dedicated image-to-video nodes that build the latent from a start image - but the T2V path, and the whole "make a clip from nothing" workflow, lives here.

The things everyone learns the hard way

• The 201-frame loop. Generate exactly 201 frames and the clip loops cleanly - Hunyuan's repetition artifact means everything past frame 200 is a repeat, and it's a feature for people making seamless loops. It was the one thing Hunyuan did natively that Wan couldn't for a year. Free loop, just set length to 201.
• It's the 1.0 node, not the 1.5 node. HunyuanVideo 1.5 is a different model with a different latent: 32 channels, 16× spatial compression, and its own empty-latent node (Empty HunyuanVideo 1.5 Latent). Feeding this 1.0-shaped latent to a 1.5 model fails or produces garbage - check which checkpoint you actually loaded. The 1.0 node even still displays "1.0" in its name precisely so you don't mix them up.
• Resolution = VRAM, brutally. The original 13B needed 45–60GB at 720p before FP8 and quantization; even at 848×480 you'll want a hefty card or GGUFs. If 1.5's 14GB floor sounds better, it is - this node is increasingly a historical artifact as the community moved to 1.5, then to Wan.

An honest word on relevance

Nobody much runs HunyuanVideo anymore - by 2026 the corpus threads had fallen to single digits a month, with the standing answer to "what happened to Hunyuan" being "nobody uses it." But its weights still circulate (FramePack builds on Hunyuan, inheriting the uncensored behavior), and if a legacy workflow or a FramePack-based graph crosses your desk, this is the node that starts it. Know what it is, use it when the graph demands it, and don't go hunting for reasons to prefer it over Wan 2.2.