Image Prompt to Video Prompt

Image-editing prompts and video prompts are not the same dialect. "Make it look like a 3D render" is a complete, sensible image instruction; a video model needs that plus motion, temporal continuity, and a role for the instruction - a system message, an instruction to the model, or a bonus directive. UC_ImageToVideoPrompt translates between the two: feed it image-flavored prompt text and it converts the language into role-aware video guidance while keeping your style, identity, and content requirements intact.

This is the small but real friction point when you build a library of image prompts and suddenly want them driving a video model. Instead of rewriting every prompt in a new dialect, you keep your existing text and let the node do the dialect shift - including appending role-appropriate guidance so the instruction lands the way the model expects.

How it works

text is your image-oriented prompt. The prompt_role dropdown picks the conversion behavior - general, system, instruction, or bonus - without changing the node's interface. Behind the scenes it's a rule-based rewrite, not an LLM call (worth knowing: no API, no key, nothing sent anywhere): it applies a set of common text replacements that shift image-editing phrasing toward motion and temporal language, then applies replacements specific to the chosen role, then appends that role's standard guidance sentence if it isn't already present.

So "make the lighting golden hour" becomes something like "… and ensure the golden-hour lighting stays consistent across frames, with the camera and subject motion behaving naturally" - the style requirement survives, the temporal instruction is added.

The inputs that matter

• text - multiline, with dynamic-prompt support, so you can drop wildcard-style syntax in if you use that.
• prompt_role - the four-way switch described above. general is the safe default; system if the text is destined for a system-message slot, instruction for a direct edit instruction, bonus for an appended enhancement block.

Single output: text, the converted string.

Installing it

ComfyUI Manager → search "ComfyUI-UtilsCollection", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart. Deps: opencv-python, typing-extensions - nothing heavy, and explicitly not an API node, so there's no account or key setup.

The honest limits: rule-based translation is clever but not magic. Highly idiosyncratic phrasing may come through with the role guidance appended and the core wording barely changed - which is fine, that's the safe direction (it preserves your content and adds the missing temporal frame). And empty input returns empty output silently, so a blank text box won't error - it'll just give you an empty prompt further down, which can be confusing mid-debug.