TextEncodeHunyuanVideo_ImageToVideo

Hunyuan Video's image-to-video is unusual under the hood, and this is the node that shows it: instead of encoding your prompt and handling the first frame as a separate concat trick, Hunyuan I2V interleaves the image's visual tokens directly into the prompt token stream. The text and the image are read together as one sequence, like giving the model a picture and a caption in the same breath. TextEncodeHunyuanVideo_ImageToVideo is the encoder for exactly that setup, and its sibling node HunyuanImageToVideo is what takes the result and builds the latent.

How it works

You feed it a clip (the Hunyuan Video text encoder), a clip_vision_output (the CLIP-Vision embedding of your start image, produced by a CLIPVisionEncode node), and a prompt. Under the hood it wraps your prompt in a Hunyuan-specific template for I2V, injects the vision model's projected image embeddings into the token sequence, and encodes the whole thing through the text encoder. The image_interleave input (default 2) controls how the image tokens and text tokens are interleaved - and its tooltip says it best: it sets "how much the image influences things vs the text prompt. Higher number means more influence from the text prompt." So a low value leans on the image, a high value leans on your words. If your I2V output keeps ignoring what you wrote, raising it is the first dial to touch; if the model keeps redrawing the image's subject into something else, lower it.

The single CONDITIONING output goes into HunyuanImageToVideo's positive input (which, worth noting, has no negative slot of its own - this is the whole conditioning side of the I2V pipeline).

How you get it

Ships with ComfyUI core. The models: the Hunyuan Video checkpoint, its text encoder, and a CLIP-Vision model for the image side - all in the standard models/ folders.

Where people get burned

Missing or wrong vision model. The node requires a real clip_vision_output; leave it unwired and there's nothing to interleave, so you get a conditioning that ignores your image entirely. Using a CLIP-Vision that wasn't trained for Hunyuan's embedding shape is the "works but drifts badly" failure. Prompt style. Hunyuan responds to detailed, photorealistic, descriptive prompts - the terse tag style from image models underperforms here, and for I2V especially the prompt should describe motion and change, not re-describe the image. The bigger picture. Be aware of where this model sits in 2026: Hunyuan got overtaken by Wan and LTX, its own I2V came late, and if you're choosing a video stack from scratch, Wan 2.2 or LTX 2.3 will take you further. But if you're running a Hunyuan I2V workflow - or you want the only open video line where image and text genuinely share one token stream - this is the node that makes it go.