Hunyuan Video T2V Prompt Generator 🐑

Text-to-video models don't want the terse tag-soup prompts you'd throw at an SDXL checkpoint. They want a paragraph: subject, motion, camera, lighting, mood. Writing those well is a skill, and this node outsources it. You give it a rough idea and an LLM connection, and it rewrites your idea into a detailed, Hunyuan-flavored video prompt.

It's part of ComfyUI-MieNodes (ComfyUI_MieNodes), MieMieeeee's utility pack - the author is a video creator, and the prompt-generator family is clearly built from real production use. It sits under 🐑 Prompt Generator.

Why you'd reach for it

HunyuanVideo was the model that made local text-to-video feel real, and its calling card was a distinctly cinematic default look. But cinematic output needs cinematic prompting - it rewards describing the shot, not just naming the subject. The KB's prompt-engineering notes make the general point that how you frame an instruction to a model is most of the result, and that goes double for video, where you're also specifying motion the model can't infer from a noun.

Be realistic about the model, though. Per the KB's Hunyuan panel, Hunyuan's local usage has fallen off hard through 2026 - Wan and LTX own the conversation now, and the standing community read is that few people still reach for Hunyuan text-to-video. If you're already committed to a Hunyuan T2V workflow, this node is a real help; if you're picking a video model from scratch in 2026, know that this one is a fading choice and the prompt style it targets is Hunyuan-specific.

How it works

Three inputs, one string out. It takes a connection to a large language model, your raw input_text, and a seed, and asks the LLM - steered by a built-in system prompt tuned for Hunyuan text-to-video - to expand your idea into a full prompt. The LLM does the heavy lifting of turning "a fox running through snow" into a paragraph with camera movement, pacing, and atmosphere. The seed makes it repeatable and gives you a reroll knob: change it to get a different expansion of the same idea.

The catch that trips people up: it needs an LLM connector, and that's a separate node. MieNodes ships a family of Set…LLMServiceConnector nodes for various providers (several with free tiers). You configure one of those with an API key, wire its output into this node's llm_service_connector, and that's what actually does the rewriting. This node is the Hunyuan-specific wrapper around it.

The inputs and outputs that matter

• llm_service_connector (LLMServiceConnector) - required. Comes from one of the pack's Set-connector nodes. No connector, no output.
• input_text (STRING, multiline) - your rough idea. A sentence is enough; the LLM expands it.
• seed (INT) - reproducibility and reroll. Same seed + same input = same prompt.

Output is hyvideo_t2v_prompt (STRING). Wire it into your text-encode / positive conditioning for the Hunyuan sampler.

Installing it

ComfyUI Manager → search ComfyUI-MieNodes → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes

then restart. The node itself is light, but using it means an LLM call, so you'll need an API key configured on a connector node. Nodes appear under 🐑 MieNodes.

Common issues & troubleshooting

"It does nothing" usually means no connector. The llm_service_connector input is required and it's easy to miss that it comes from a separate Set-connector node you have to add and configure with a key. Wire that up first.

API errors. A bad or empty API key, an out-of-credit account, or a model id the provider doesn't recognize all surface as connector failures. Check the key and the selected model on the connector node, not on this one.

The prompt is great but the video ignores it. That's a known Hunyuan trait, not a bug in the generator - the KB records Hunyuan's prompt adherence as its weak spot, especially on image-to-video. A better-written prompt helps at the margin; it can't make the model follow instructions it structurally doesn't. For a starting image plus motion, use the I2V variant instead.

Same output every run. Change the seed. It's deterministic by design, so a fixed seed reproduces the same expansion.