🎞️SVD Aspect Ratio

SVD works best when you tell it the shape of the video up front instead of letting a random latent pick for you. 🎞️SVD Aspect Ratio is the node from the ComfyUI-SVD-ZHO pack that does exactly that - it hands you the three aspect ratios the old stablevideo.com web app offered, plus a batch count, and produces the empty latent those dimensions need.

It's about as simple as a node gets, but simple here is a feature: it removes a decision you'd otherwise make inside a generic Empty Latent Image node, and it hardcodes the exact sizes the pack's SVD workflow expects.

How it works

Under the hood it's a tiny torch call. It builds a zeroed latent tensor of shape [batch_size, 4, H/8, W/8], where the /8 is the VAE downscale factor. The three presets:

• 16:9 → 1368×768
• 1:1 → 1024×1024
• 9:16 → 768×1368

Notice 1368 divides cleanly by 8 (1368 = 8×171), which is exactly what a legal ComfyUI latent needs - no off-by-one shape errors. That's the whole trick, and it's the kind of detail that bites you if you set dimensions by hand.

Inputs

• Aspect_Ratio - an enum, and only the three choices above. That's it.
• batch_size - 1 to 64, default 1. How many identical empty latents to make.

Output

A single LATENT, named latent. In the pack's text-to-video flow this is your canvas: generate a first frame at this size (the README chains Styler → Aspect Ratio → T2I Chooser), then SVD animates from it, inheriting the resolution. For image-to-video you'll usually skip this node entirely and load an image - SVD takes its size from the source frame instead.

Install

Same pack, same routine. ComfyUI Manager → search ComfyUI-SVD-ZHO → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-SVD-ZHO

then restart. No requirements.txt, no model downloads - it's pure torch, which you already have. Find it under Zho模块组/🎞️SVD.

Gotchas

Two things to know. First, batch_size above 1 gives you N identical empty latents - useful if you're sampling multiple candidate first frames in one pass, pure waste otherwise. Second, the three ratios are hardcoded; if you want 4:3 or an ultrawide 2.39:1, this node can't do it and a core Empty Latent Image node is the better tool. This one is a convenience lock-in to the pack's sizes, not a general-purpose tool - which is fine, as long as you know that going in. The pack is WIP (v0.9) and SVD itself is a legacy model in 2026, so treat this as a small helper for a specific old workflow rather than something you'll build everything on.