EmptyCosmosLatentVideo

EmptyCosmosLatentVideo builds the blank video latent that NVIDIA's Cosmos video world models denoise from. It's the txt2video starting block for the Cosmos family: where the text-to-image Cosmos models start from a flat 2D blank, this node creates the 5D tensor [batch_size, 16, frames, height, width] - sixteen channels, with the frame count and both spatial dimensions compressed by a factor of 8. If you've only touched Cosmos through Anima (the anime model built on the Cosmos-Predict2 2B base), note this is the video side of the family, not that image model.

Four inputs, all self-explanatory until you hit the frame math:

• width (default 1280, step 16) and height (default 704, step 16) - the output clip's pixel size.
• length (default 121, step 8) - the number of frames. This is where it gets fiddly: the node compresses temporally by 8×, so 121 frames becomes a latent of only ~16 time-steps, and the slider enforces that step of 8.
• batch_size (default 1, up to 4096).

Output is a LATENT that goes into the sampler. Simple enough - but the context matters more than the inputs.

Where it fits, and why it's niche

Cosmos is a world-model family for robotics and simulation that this community mostly ignored - then Anima made the 2B Text2Image variant famous as an anime backbone. The video world models stayed what they always were: heavy, slow, and aimed elsewhere. A 24GB card at default settings runs a clip in roughly twenty minutes, which the community recorded at Cosmos's ComfyUI debut and nobody since has called fast. The realistic take: you'll reach for this node if you're experimenting with Cosmos Video2World checkpoints for actual world-model work, not for casual clip generation. The pairing node to know is CosmosPredict2ImageToVideoLatent on the conditioning side, which builds a masked latent from start/end images; this node is the pure txt2video blank without those constraints.

Gotchas

The 8× temporal compression is not a suggestion - length steps in 8 and the latent rounds down, so a clip length that isn't a multiple of 8 decodes at a slightly shorter frame count than you asked for. And remember this latent format is Cosmos-specific: 16 channels at 8× spatial and 8× temporal compression is not the same shape as Wan or Hunyuan video latents, and swapping one blank into the wrong model's workflow produces nonsense, not a slightly-off clip. If you're chasing the anime image models, this is the wrong node entirely - the image side starts from the flat EmptyLatentImage-style blanks, not this 5D tensor.

It ships with ComfyUI core; the Cosmos checkpoint and its VAE are the downloads that matter. Worth knowing it exists, worth knowing it's slow, and worth knowing it's aimed at a very different audience than the rest of this category.