Hunyuan Video I2V Prompt Generator 🐑

Image-to-video prompting has a trap: people re-describe the picture they already handed the model. But the model can see the starting frame - what it needs from you is what should happen. This node is built around that distinction. You describe the source image and your intent, and an LLM writes a Hunyuan image-to-video prompt focused on motion and change rather than restating what's already on screen.

It's part of ComfyUI-MieNodes (ComfyUI_MieNodes), MieMieeeee's utility pack, under 🐑 Prompt Generator.

Why you'd reach for it

The whole point of I2V is that the first frame is given. So a good I2V prompt spends its words on the things the still image can't express: camera push, a turn of the head, wind in the trees, a shift in light over time. The KB's prompt-engineering doc frames prompting as steering a model toward what you want - and for I2V, "what you want" is the delta from the starting frame, which is a genuinely different skill from writing a text-to-video prompt.

This node bakes that framing in. It takes a separate description of the image and your input, so the LLM has both the context of where the clip starts and your instruction for where it goes. One honest caveat from the Hunyuan KB panel: image-to-video is actually Hunyuan's stronger mode - the community verdict is that it "shines in img2vid" - but even there prompt adherence is shaky. Good prompting is the lever you have; use it, but keep expectations calibrated.

How it works

Same engine as its text-to-video sibling, with one extra input. It takes an LLM connector, an image_description, your input_text, and a seed, and asks the LLM (steered by a Hunyuan-I2V system prompt) to compose a prompt that starts from the described image and adds the motion you asked for. The image_description is the piece that makes it I2V-aware - fill it in yourself, or feed it the output of a vision/caption node so the description is automatic. The seed gives you repeatable results and a reroll.

As with all these generators, the LLM lives in a separate connector node from the same pack: you configure a Set…LLMServiceConnector with an API key and wire it in. This node is the Hunyuan-I2V-specific wrapper.

The inputs and outputs that matter

llm_service_connector (LLMServiceConnector) - required; from a Set-connector node.
image_description (STRING, multiline) - what's in the starting frame. Type it, or wire in a caption node's output. This is what separates it from the T2V version.
input_text (STRING, multiline) - your instruction: the motion, the change, the shot you want.
seed (INT) - reproducibility and reroll.

Output is hyvideo_i2v_prompt (STRING). Feed it into the positive conditioning of your Hunyuan I2V 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. Using it means an LLM call, so a connector node with a valid API key is required. Nodes appear under 🐑 MieNodes.

Common issues & troubleshooting

Empty image_description wastes the node. If you leave it blank you've essentially got a worse text-to-video generator. Describe the frame - or auto-fill it from a caption node - so the LLM knows what it's animating.

No output. Almost always the missing or misconfigured llm_service_connector. It's a separate node with the API key; add and configure it first.

The prompt describes the image instead of the motion. If the LLM keeps restating the still, put the change front and center in input_text ("she turns to face the camera, slow dolly in") and keep image_description factual and short. The clearer the split between "what's there" and "what happens," the better the rewrite.

Provider/API errors surface from the connector, not here - check the key, credits, and model id on the Set-connector node.