Frame Transition Prompt Generator 🐑

First-frame / last-frame video generation is only as good as the prompt describing how you get from A to B. This node writes that prompt. You describe the starting image and the ending image, and an LLM composes a motion/transition prompt - the camera move, the action, the change - aimed at Wan-family video models. It's a small, specific tool for a specific pain: turning "start here, end there" into words a video model can act on.

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

Why you'd use it

If you've done start/end-frame video, you know the awkward middle step. You have two clear stills. What you don't have is a good sentence describing the transition between them - the motion, the timing, the camera behavior - and video models care a lot about that sentence. Writing it well, repeatedly, for different frame pairs, is exactly the kind of grind worth automating.

This node hands the job to an LLM. Feed it a description of the start frame and a description of the end frame, and it reasons about what has to happen in between and writes it up as a transition prompt. That's the LLM-assisted prompting pattern the KB's prompt-engineering doc documents as now-standard: it's twentyfold more common than it was in 2023, and it fits video especially well because the "what moves and how" description is tedious to author but easy for a language model to draft.

The model input currently offers wan - the prompt is tuned for how Wan-family models like transition instructions phrased. So this is aimed at Wan I2V / first-last-frame workflows specifically, not a generic video prompt writer.

The inputs and outputs that matter

All required:

• llm_service_connector - from one of the pack's Set...Connector nodes; the LLM doing the writing.
• start_image_description - multiline text describing your first frame. As with the other generators, you can auto-fill this with a captioner instead of typing it.
• end_image_description - multiline text describing your last frame.
• model - the target family; wan.
• seed - reroll for a different phrasing of the same transition.

Output is transition_prompt (STRING) - wire it into the positive-prompt/conditioning input of your Wan first-last-frame video workflow.

Installing it

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

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

then restart. Nothing to download for the node. It needs an LLM connector (with a key), and it produces only the prompt - the Wan video model and its start/end-frame plumbing live elsewhere in your graph. Nodes appear under the 🐑 MieNodes menu.

Common issues

The transition ignores half of what you wanted. The output is only as good as the two descriptions. If the start and end descriptions are thin, the LLM has nothing to reason a motion out of. Describe both frames concretely - what's in them, and what's different between them.

It writes a scene, not a motion. If the result reads like a static description instead of a transition, make the difference between your two frames explicit in the descriptions. The node infers movement from the delta between start and end; spell that delta out.

Wrong-model phrasing. The prompt is shaped for Wan. If you're driving a different video model, the wording may not land the same way - treat the output as a strong draft you can hand-tweak, not gospel.

Errors before anything generates. That's the upstream connector (key, endpoint, or model id), not this node.