Video Size (Local Models)

Video models are the pickiest things you'll ever run locally, and it's not a vibe thing - it's mechanical. Wan, Hunyuan Video, and CogVideoX are each trained at specific resolutions, and deviating from them means weird crops, resolution-specific VAE mismatches, or just garbage motion. Video Size (Local Models) exists so you don't have to remember which model wants what. It's the video sibling of the pack's image-size nodes and ships in the Open Creative Studio Nodes suite that accompanies the OCS workflow (formerly AP Workflow).

How it works

One dropdown, a handful of preset resolutions, each labeled with the model it belongs to:

• 1360x768 [CogVideoX 1.5]
• 1280x720 [WanVideo 2.2, Hunyuan Video]
• 960x544 [Hunyuan Video]
• 854x480 [WanVideo 2.2]
• 720x480 [CogVideoX 1.5]

Pick one and the node overrides your custom width and height, then hands back the dimensions plus a blank latent at H//8 × W//8. The //8 reflects how these latent diffusion models encode pixels. Worth a small note that the README still says "WanVideo 2.1", but the node ships presets for 2.2 - trust the node, the README just rotted.

Inputs that matter

• Aspect Ratio - the preset list. This is your entire interaction with the node 90% of the time.
• Custom Width / Custom Height - used when you set Aspect Ratio to custom (64–8192). Handy when you're upscaling a video or matching an input clip's dimensions.
• Batch Size - 1 to 64.

Outputs and the honest caveat

Outputs are aspect_ratio (a 1280x720 string), video_width, video_height, batch_size, and video_latent. Here's the trap: the latent output is a 4-channel zeros tensor - a 4D canvas. Real video latents are 5D, with a frames dimension, because a video latent covers multiple frames at once. So don't wire video_latent into a video sampler and expect it to work. The outputs that actually do work across models are the three INTs and the aspect string, which you feed into whatever your workflow uses to build its real empty latent - or into an EmptyLatentVideo if that's your model's idiom.

This matters more than it sounds: Wan's 720p and 480p variants are effectively different products, and Hunyuan has its own resolution-specific VAE. Getting the resolution wrong here isn't a cosmetic problem, it's a "why is my video 10 frames of noise" problem.

Install

ComfyUI Manager → "Open Creative Studio Nodes" → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/alessandroperilli/OCS_Nodes

No extra dependencies, no bundled models. The pack only relies on ComfyUI's own environment.

Gotchas

• The latent output is a dummy. Treat it as a sizing hint, not something to wire into a sampler - use the INTs.
• Presets only cover the current gen. If a new video model lands without an OCS preset, custom is right there; you just have to know its native resolution yourself.
• Batch size on video is brutal. Video batches multiply VRAM fast - don't treat 64 as an invitation.