CosmosPredict2ImageToVideoLatent

Cosmos-Predict2 is NVIDIA's video-prediction "world model" - built for Physical AI and robotics, which is exactly why nobody here adopted it as an art model, even though ComfyUI support landed within a week of its June 2025 release. This node is its image-to-video latent prep: the thing that turns your start frame (and optional end frame) into the noise target the Cosmos sampler denoises.

Inputs

• vae - the Cosmos VAE. Required.
• width / height / length / batch_size - video dimensions. Defaults are 848×480, 93 frames.
• start_image - optional. The first frame of your clip.
• end_image - optional. A target final frame. This is the "Predict2" trick: the model isn't just predicting forward from a start, it can be told where it's supposed to land, and it interpolates between them.

How it works

The node allocates the latent: [1, 16, T, H/8, W/8] - 16 channels, temporal compressed by 4, spatial by 8. If you supply start and/or end images, it VAE-encodes them into the first and last latent slots and builds a mask so the sampler knows those frames are fixed and everything between is free to predict. No images, and you just get a blank latent - which makes the node double as a text-to-video latent builder for the same checkpoint.

The workflow

Wire the latent into a KSampler along with the Cosmos model - the ModelMergeCosmosPredict2_2B / _14B nodes handle the model side, and there's a text encoder for the prompt. This is the node that sits between your images and the sampler, and unlike an SD-style empty latent it's dimensionally exact for Cosmos, so don't substitute an EmptyLatentImage.

The honest context

Two things to know before you build this. First, the model is a world model, not an art model. It predicts plausible video - great for simulation-adjacent work, not your default for cinematic clips. Second, and this is the one that surprises people: the Predict2 repository was archived in December 2025 in favor of Cosmos-Predict2.5, so the base line is frozen. The node still works, the weights still load, and the license (NVIDIA Open Model License, genuinely permissive) still holds - but don't expect upstream improvements.

The reason this family got interesting anyway is downstream: Cosmos-Predict2's 2B text-to-image became the backbone of Anima, the anime model. That's a different pipeline, though. If your goal is the I2V node on this page, you're using Cosmos for what it is: a competent, fast, world-model video generator that the community tried, shrugged at, and moved on from. It ships with ComfyUI core - no install.