(Deno) Bernini Prompt Guide

Bernini is ByteDance's video editing model - Wan 2.2's weights reworked into a maskless instruction-following machine - and the single biggest thing to know about prompting it is that it does not prompt like a caption model. You don't write "a man in a red shirt walking through a market." You write instructions: "Replace the jacket with the shirt from image0. Keep camera motion, face, lighting, and background unchanged." The (Deno) Bernini Prompt Guide is a KJ-style helper that bakes that instruction format in, so you stop fighting the model and start directing it.

What it does

It's a text-encoding node: CLIP in, positive and negative conditioning out. The task selector (task_type) picks the system-prompt mode - Text to Video, Image to Video, Reference Video Edit, and friends - and the active instruction prefix is shown right on the node so you always know what's being prepended. In reference modes, a reference_prompt_helper toggle adds the image0 / image1 / image2 naming hints internally, which is exactly how Bernini's multi-reference prompting works: you index the reference images by name and tell the model what to keep and what to swap.

The negative side is collapsible, and negative_preset fills the box with the official Wan 2.2 negative - the long Chinese-language list - which you can then edit directly. What you see in the box is what gets encoded, so if you want a leaner negative, just delete and type.

Inputs and outputs that matter

• clip - the Wan/Bernini text encoder. Wire in the CLIP from your Bernini/Wan 2.2 load.
• task_type - the mode. Reference Video Edit is the default and the one you'll live in for v2v work.
• positive_prompt - your instruction, written like a clear chatbot command.
• negative_preset / negative_prompt - preset fills, box is authoritative.
• Outputs: positive and negative CONDITIONING for the sampler.

The caveat you have to know

This node prepares text conditioning only. Bernini's visual conditioning (the reference-image context latents) still needs a ComfyUI/KJ backend that supports them - and as of the current docs that backend support is a draft ComfyUI PR. The README is blunt: use tools/DENO_Bernini_Preview_Backend_Update.bat only on a copied/test portable ComfyUI folder, never your main install. In practice that means this node slots into the Kijai Bernini workflow (the renderer-only Bernini-R path nearly everyone actually runs), and it earns its keep there by making the prompt half of that workflow one tidy node instead of hand-assembled text.

Install is the standard pack path:

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
# restart ComfyUI

The gotcha beginners hit is expecting this node alone to give them a working Bernini pipeline. It won't - you still need the Bernini model files, the ComfyUI support, and the compute budget. Bernini editing runs roughly double a Wan 2.2 generation because the source video is part of the model input. The node just makes the prompt side correct and repeatable, which is a real thing. Once you've had one output where the model actually swapped the jacket and kept everything else, you'll understand why the format matters.