HunyuanVideo Text Encoder

This is the node that turns your prompt into the positive and negative conditioning HunyuanVideo 1.5 needs for text-to-video. Feed it the Qwen2.5-VL encoder and your text; it hands back both a positive and a negative in one shot, ready for the KSampler. The convenience twist is the built-in template system: instead of hand-writing a system prompt that describes cinematography, motion, and pacing, you pick from ~40 video templates (cinematic, documentary, animation, action, aerial, and so on) and let it frame your prompt appropriately.

Because HunyuanVideo's encoder is a language model, the system prompt genuinely steers the look of the motion, not just the subject. The templates are a shortcut to prompts the author already tuned for particular video styles - and you can layer your own tweaks on top without throwing the whole template away.

How it works

Your prompt rides the Qwen2.5-VL chat template, with the chosen video template supplying the system-prompt framing. The node emits paired positive and negative conditioning - the negative comes pre-populated with a sensible default ("low quality, blurry, distorted, artifacts, watermark, text, logo") so you're not starting from nothing. Two escape hatches keep it flexible: additional_instructions layers modifications on top of a template (keep the cinematic frame, but add "slow dolly-in"), while custom_system_prompt throws the presets out and lets you write the system prompt entirely yourself.

The output shape - positive and negative together - is designed to drop straight into a standard KSampler (or SamplerCustomAdvanced), which is the native ComfyUI HunyuanVideo path this pack rides on.

The inputs and outputs that matter

• clip (required) - the Qwen2.5-VL encoder from HunyuanVideoCLIPLoader.
• text - your prompt: describe the scene and the motion.
• template_preset (default none) - one of ~40 video-style templates. The fastest way to get cinematic-looking output without writing a system prompt.
• additional_instructions - modifications layered onto the chosen template. Use this for "same style, but change X."
• custom_system_prompt - full manual control; overrides the preset when you want your own framing.
• negative_prompt - comes with a usable default; edit it to exclude specific artifacts.

Outputs: positive and negative conditioning (both into your KSampler), plus debug_output.

How to install it

ComfyUI Manager: search ComfyUI-QwenImageWanBridge, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-QwenImageWanBridge

then restart. Needs the Qwen2.5-VL-7B encoder (via HunyuanVideoCLIPLoader) plus the HunyuanVideo 1.5 diffusion model and VAE. The example flow is HunyuanVideoCLIPLoader → HunyuanVideoTextEncoder → KSampler → VAEDecode.

Common issues & troubleshooting

The motion is generic or static. Lean on the templates - cinematic, action, and the motion-focused presets frame the prompt with the cinematography language the model responds to. A bare subject description without motion guidance tends to produce a near-still clip. Describe what moves.

Your prompt is huge and results drift. These LLM encoders lose the thread past ~75-100 effective tokens. State a few strong elements and the key action; a 300-word shot description works against you.

Prompt weighting / negative embeddings don't work. Same as everywhere in this pack - that's CLIP-era syntax the Qwen2.5-VL path never implemented. Steer with the template, additional_instructions, and a plain-language negative_prompt instead.