🎞️ Grid Storyboard Prompt

Image models can draw a 2x2 or 3x3 grid in a single image, but getting them to treat each panel as a different moment of the same story - same characters, same style, consistent lighting across the whole sheet - is where hand-written prompts fall apart. Pix_GridStoryboardPrompt assembles exactly that prompt for you: it takes your style, scenes, subjects and per-panel shot descriptions and merges them into one structured, spatially-annotated text aimed at grid-image generation models like Flux, SDXL, or NanoBanana Pro.

Five text inputs plus two menus, and the output is a single prompt string. The inputs: style (the overall aesthetic), scene (environment/lighting details), subject (characters, outfits, appearance), and shots (what happens in each panel - the most important field), plus grid (2x2, 3x3, 4x4) and ratio (nine aspect ratios from 21:9 to 9:21).

The clever part is the data parsing. Each text field accepts plain text, a JSON array (["...", "..."]), or an array of objects - shots reads the prompt key from object entries, the others read content. That means you can feed shots straight from a structured source, including an LLM that writes storyboard JSON, and the node will line each shot up with a spatial position. For a 2x2 grid it names positions 左上 (top-left), 右上, 左下, 右下; for 3x3 and 4x4 it uses "第 1 行，第 2 列" row/column notation. Every panel line is prefixed with its position, and the header explicitly instructs the model to keep style, characters, lighting and palette consistent across all panels - which is the single most important instruction for grids, since models will happily drift between panels otherwise.

Output sections are composed conditionally: only the blocks you filled in appear, so an empty scene field doesn't produce a paragraph of nothing. The whole thing comes out in Chinese - that's deliberate, it's built for Chinese-ecosystem models and prompt styles, and it works fine with them. If your base model responds better to English prompts, you'll want to run the output through a translate node before use.

This node sits in the pack's AIGV (AI-generated video/storyboard) corner, next to the Pix_StoryboardPreviewer, and it's meant to feed a grid-sampling workflow. The workflow examples the pack ships show it driving LLM-assisted storyboard generation - an LLM produces the shot JSON, this node formats it, and the sampler draws the grid.

Install is the standard PixNodes path - ComfyUI Manager, search PixNodes, restart - or:

cd ComfyUI/custom_nodes
git clone https://github.com/pixixai/Comfyui-PixNodes
cd Comfyui-PixNodes
pip install -r requirements.txt

then restart. Just openai and Requests; no models.

Honest take: this is a niche but genuinely useful node for anyone doing character-consistent multi-panel grids. The consistency instruction in the header is the whole game - grids fail on consistency, and this node at least makes your prompt say the right thing. Whether the model obeys is still sampling roulette, but you're no longer starting from a blank prompt.