ARVideoI2V - ComfyUI Built-in Node Documentation

Overview

This node prepares an image-to-video generation setup for AR (Auto-Regressive) video models that use Causal Forcing or Self-Forcing. It encodes a starting image into latent space with a VAE and stores it in the model’s transformer options, so the video sampling process can seed the KV cache before denoising. It uses the same text-to-video model checkpoint, so no separate image-to-video architecture is needed.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The AR video model to be used for generation.	MODEL	Yes	-
`vae`	The VAE model used to encode the starting image into latent space.	VAE	Yes	-
`start_image`	The initial image that will serve as the first frame of the generated video. Only the first image in the input batch is used, and only its RGB channels are encoded.	IMAGE	Yes	-
`width`	The width of the generated video frames (default: 832).	INT	Yes	16 to 8192 (step: 16)
`height`	The height of the generated video frames (default: 480).	INT	Yes	16 to 8192 (step: 16)
`length`	The total number of frames in the generated video (default: 81).	INT	Yes	1 to 1024 (step: 4)
`batch_size`	The number of video sequences to generate in a single batch (default: 1).	INT	Yes	1 to 64

Note: The start image is resized to the specified width and height before being encoded. The latent temporal dimension is calculated as ((length - 1) // 4) + 1, and the latent spatial dimensions are height / 8 and width / 8.

Outputs

Output Name	Description	Data Type
`MODEL`	The cloned model with the encoded start image stored in its transformer options (`ar_config.initial_latent`), which the sampler uses to seed the KV cache before denoising.	MODEL
`LATENT`	A zero-filled latent tensor with shape `[batch_size, 16, lat_t, height // 8, width // 8]`, where `lat_t = ((length - 1) // 4) + 1`.	LATENT

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 984834951b9d5a22aef51c85a5019fd8ba58cdb2d6fff235371ed29f316896d8

​Overview

​Inputs

​Outputs

Overview

Inputs

Outputs