HunyuanVideo Empty Text Embeds

This is about as simple as a node gets: no inputs at all, one output. It hands the sampler an empty hyvid_embeds - a valid, unconditional conditioning object - without you needing to load the LLM text encoder and run an actual prompt through it just to get an empty one.

The description spells out why you'd want that: it's "to avoid having to encode prompts for inverse sampling." Inverse sampling here refers to the wrapper's RF-Inversion support (visible in the pack's source as a dedicated module) - a video-editing technique that works by running the diffusion process in reverse to invert a real video into its latent noise, then re-generating with modified conditioning. Some steps of that pipeline genuinely don't need a text prompt at all, and loading an 8B-parameter LLM purely to generate a null embedding is wasted time and VRAM you don't have to spend.

When you'd actually use this

If you're doing a normal text-to-video or image-to-video generation, you don't need this node - you want HyVideoI2VEncode or the text-encode path with an actual prompt. This node is specifically for workflows where an unconditional embed is genuinely all that's needed for a given stage: inversion/editing pipelines, or any experimental workflow where you deliberately want "no text conditioning" as a baseline to compare against.

The inputs and outputs

Nothing to configure - required and optional inputs are both empty. Output: hyvid_embeds, wired the same place any other encode node's output goes: into HyVideoSampler.

How to install it

Ships with the pack - no separate download for this node.

• ComfyUI Manager - search ComfyUI-HunyuanVideoWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-HunyuanVideoWrapper, then pip install -r ComfyUI-HunyuanVideoWrapper/requirements.txt, restart.

You don't need the LLM text encoder for this node specifically - that's the entire point - but you'll still need the diffusion model and VAE loaded for the rest of the pipeline to run.

Common issues & troubleshooting

There's genuinely not much to troubleshoot on a zero-input node like this - if it's producing an empty embed, it's working. The one real mistake is reaching for it by accident: if you wire this into a sampler expecting your actual prompt to matter, you'll get generic, unguided output and no obvious error telling you why. If a generation looks like it's ignoring your prompt entirely, check you didn't accidentally leave this node connected instead of a real text-encode node from an earlier experiment.