HyVideo TextEmbed Bridge

HunyuanVideo normally reads your prompt through an 8B-parameter LLM (see DownloadAndLoadHyVideoTextEncoder), which is expensive to load and run. This node is kijai's answer to "does it have to be that way": it takes native ComfyUI CONDITIONING - from whatever CLIP or T5 text-encode node you already have in your graph - and reformats it into the HYVIDEMBEDS type the sampler expects. No LLM required.

The author's own framing, straight from the README: it gives "somewhat different results" and "can't be considered as original implementation wrapper anymore." That's an honest warning, not a knock - this is a legitimate way to save VRAM and load time if you're willing to accept the model isn't being fed conditioning quite the way it was trained to expect. Worth trying if the LLM text encoder is the thing standing between you and this pack actually running on your card.

The inputs that matter

• positive - a native ComfyUI CONDITIONING output, from any CLIP/T5 encode node in your graph.
• cfg (default 1) - the classifier-free guidance scale for this bridged conditioning.
• start_percent / end_percent (0 to 1) - the tooltip is direct about this: they mark the slice of the schedule where this CFG value applies, and the rest of the steps fall back to the model's baked-in guidance_embeds. HunyuanVideo is a guidance-distilled model, so it doesn't need real CFG applied at every step the way an older model would.
• batched_cfg - runs the conditional and unconditional pass as a batch instead of sequentially. Faster, costs more VRAM.
• use_cfg_zero_star (default on here) - a refinement to classifier-free guidance aimed at fixing an early-step artifact; on by default in this node, which is a signal the author considers it a safe default.
• negative - optional CONDITIONING, for when you actually want a negative prompt to matter (it only does anything during the percentage window where cfg is active).

Output: hyvid_embeds, straight into HyVideoSampler.

How to install it

Ships with the pack.

• 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.

The whole point of this node is that you don't need the LLM text encoder - you still need the diffusion model and VAE (manual download from Kijai's HunyuanVideo_comfy repo on HuggingFace), but you can skip DownloadAndLoadHyVideoTextEncoder entirely and wire in ComfyUI's own CLIP loader instead.

Common issues & troubleshooting

Results look noticeably different from the "real" pipeline, or prompt adherence feels off. That's expected and stated by the author directly - this is a deliberately different conditioning path, not a drop-in replacement. If fidelity to the original model's behavior matters for your use case, go back to the LLM-based encoder nodes instead.

Negative prompt seems to do nothing. Check start_percent/end_percent actually cover part of the schedule - if the window is zero-width or your cfg is left at 1 with a narrow window, the negative prompt barely gets a chance to influence anything before the model falls back to its embedded guidance.

Generation is slower than you expected from "skipping the LLM." batched_cfg and a wide CFG window both cost real compute - this node saves you the LLM's load and forward-pass cost, not the cost of running true classifier-free guidance if you've turned that on for most of the schedule.