NL Bernini Prompt Enhancer

The name is slightly misleading, so let's clear it up first: NL Bernini Prompt Enhancer doesn't enhance anything itself, and it doesn't call any API. What it does is build the prompt strings that a downstream GPT/API node should send for Bernini's prompt-enhancement modes. Bernini - ByteDance's video generation and editing model built on Wan 2.2 - is prompt-based in a way that rewards structured, task-specific prompting: its editing modes want to know whether you're doing video-to-video editing, multi-reference editing, image-to-video, and so on. This node hands you a correctly-structured template for whichever of those tasks you're doing.

Think of it as the form you fill out so your LLM gets the right instructions. You pick a mode, you type your artist instruction, and the node emits the system prompt, user prompt, and metadata that a GPT-style node needs. The README is explicit that it builds these strings without calling an API or storing keys - it's pure string assembly with an opinion.

The inputs and outputs that matter

• mode - the task template. The list runs from Video Edit (v2v) through Multi-Visual Video Edit (mv2v), Reference Video Edit (rv2v), Image Edit (i2i), Image To Video (i2v), Text To Video (t2v), and on. Twelve modes total; the node's context menu has a "Bernini mode guide" entry that describes each one, which is worth reading before you pick.
• user_prompt - your raw artist instruction, the one thing you actually write.
• reference_image_count / source_video_frame_count - how many references or source frames you'll attach separately to the API node, so the generated prompt knows to expect them.

Outputs: full_prompt (connect this to a single-prompt API node), or system_prompt + user_prompt separately (for API nodes that support chat roles), plus task_type and response_format for the downstream call.

How it fits the Bernini world

Bernini's prompting is instruction-style, not caption-style - you tell it what to do with the references in plain language, and the multi-reference modes expect the references to be indexed ("the man from image0…"). This node is essentially a structured wrapper around that reality. For the solo user running Bernini through a ComfyUI API node, it saves you from hand-writing system prompts that are easy to get subtly wrong. For anyone else it's a template library in node form.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/NOLABEL-VFX/ComfyUI-NL_Nodes

or ComfyUI Manager → "ComfyUI-NL_Nodes", restart. No API keys, no models, no dependencies beyond the base set - which is the nice thing about a node that just formats strings. The catch is that it's useless without a downstream GPT/API node to feed, so make sure your Bernini API workflow is already working before you add this as an upgrade.