🔹 CYH Latent | Video Aspect Ratio

Video generation models are picky about frame dimensions in a way image models just aren't - generate too big and you blow your VRAM on a latent that's 16× larger than a still, too small and you get mush. This node is a dropdown of sane video resolutions - 1080p 16:9, 21:9 cinema, 4:3, and 9:16 portrait - that produce an empty latent ready for your video sampler.

It's the video entry in the Chye ComfyUI Toolset latent family, sharing the same code skeleton as the Flux/SDXL/phone/social variants. The default orientation is Landscape, because that's what 16:9, 21:9, and 4:3 want - though the 9:16 preset is obviously meant to be flipped.

How it works

Same pattern as the rest of the pack: the dropdown label contains the resolution, the node parses the ratio, applies orientation and multiplier, rounds to a multiple of 32, and creates a zeroed tensor:

latent = torch.zeros([batch_size, 4, final_height // 8, final_width // 8])

One LATENT out, straight into your video model's sampler. Keep in mind it's one latent image - it produces a blank frame, not a latent video. Your video node (LTX-Video, Wan, AnimateDiff) usually wants a single latent frame and generates the rest itself.

The presets

• 16:9 (Standard) - 1920×1080
• 21:9 (Cinema) - 1920×810
• 4:3 (Traditional) - 1024×768
• 9:16 (Portrait) - 1080×1920

Here's the reality check: most local video models aren't going to sample a full 1920×1080 in one pass without a serious GPU. The 4:3 1024×768 and a downscaled 16:9 are the workhorses for local generation - LTX-Video-class models live comfortably around 512–768px frames. The multiplier (0.1–10.0) is your friend: 16:9 at 0.5 gives you 960×544, a very comfortable video-model size. 1920×1080 full-res is aspirational on most rigs.

The inputs

Same four as the other latent nodes:

• aspect_ratio - the four presets above.
• orientation - Portrait/Landscape, default Landscape.
• multiplier - 0.1–10.0; this is where you right-size for your model and VRAM.
• batch_size - 1–64. This is genuinely useful for video: some pipelines want a batch of starting frames, and this produces a stack of identical blank latents.

One honest gripe: the README promises "smart defaults based on the aspect ratio" - i.e., 9:16 auto-flipping to Portrait. The code doesn't do that. The orientation dropdown has a fixed default (Landscape) and it won't change when you switch the preset, so set Portrait yourself for 9:16. Minor, but the README oversells it.

Install

ComfyUI Manager (search "Chye ComfyUI Toolset"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/chyer/Chye-ComfyUI-Toolset
cd Chye-ComfyUI-Toolset
pip install -r requirements.txt

Restart after. Notable deps: scipy, opencv-python; torch/numpy come with ComfyUI. Zip installs: add .git/.cnr-id with content Chye-ComfyUI-Toolset as the README instructs.

Honest verdict

If you feed video models from a text-to-video UI preset anyway, this is redundant. If you build video workflows by hand, having the standard frame sizes one click away - and batch-of-blanks on demand - is a small but real convenience. Just remember to scale down, not up.