首尾帧过渡提示词生成 (Comfly/T8)

A bunch of video models - Wan, LTX, Kling-style first-last-frame workflows - want a prompt that describes the whole journey from frame A to frame B, not two separate scene descriptions. Writing that by hand is tedious and the result is usually stiff ("a room... then a brighter room"). This node takes your first frame and last frame as images, ships them to a vision LLM on a Comfly or T8 mirror API, and hands you back one continuous English prompt describing the transition: "a person smoothly walking from the left side of a room to the right side, transitioning from dim to bright lighting, the camera follows the movement…". You feed that straight into the video generator's prompt slot.

It's one of the "Comfly/T8" family inside ComfyUI_LLM_Banana, Ken Chen's grab-bag multimodal pack. Like every node in that family, it doesn't run a model locally - it's an HTTP client with an API key - the external-api-nodes KB doc has the full frame on why that category exists and its tradeoffs. No GPU needed, but your frames leave the machine.

How it works

Under the hood it posts both frames as base64 to {base_url}/chat/completions with Authorization: Bearer, using the built-in system prompt (editable, Chinese comments included) that tells the model to analyze the start state, the end state, and everything that changes between them, then emit a comma-separated, third-person prompt with quality words tacked on the end. The default model is gpt-4-v - you want something with vision here, not a text-only chat model.

The inputs that matter

• first_frame / last_frame - the two IMAGE wires from your Load Image nodes. Both are required.
• api_provider - pick your mirror: Comfly, Comfly-HK/US, or the T8 variants. The base URL and key come from ChatFly_config.json (prompt_api_providers) unless you override them.
• api_key / base_url - leave blank to fall back to the config file; fill them to override.
• model - the dropdown is built from the providers in the config, so yours may differ from the default list.
• system_prompt - you can swap the built-in expert prompt for your own if you want a different flavor.
• user_requirement - extra instructions in plain language ("keep the camera static").

Output is a single transition_prompt STRING. Wire it into whatever video node you use.

Install

ComfyUI Manager, search ComfyUI_LLM_Banana. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_LLM_Banana
cd ComfyUI_LLM_Banana
pip install -r requirements.txt

then restart ComfyUI. The pack pulls a chunky dependency list (google-genai, transformers, opencv-python, openai, httpx…), so give the install a minute.

Where people get burned

The node is polite about failure: if no API key or base URL resolves, it returns an error as the output string instead of throwing, so you'll see "未检测到API Key或Base URL…" in your prompt field rather than a red node. That's the signal to check ChatFly_config.json or fill the node fields. Also: pick a vision model. Ask gpt-4 (text-only) to describe two images and you get nothing useful back, just a refusal or filler. And remember these mirror services charge per call - a video project generating a transition prompt every frame pair adds up fast.