VSmartPrompt

What it is

The {a|b} bracket trick is older than most ComfyUI users: A1111-era dynamic prompts let you write one prompt that rolls a different combination every run, and people used it to batch 32 images of the same character across hair colors, weather, and lighting with zero retyping. VSmartPrompt is that idea, rebuilt for ComfyUI with a syntax-highlighted text editor bolted on. It's the flagship of the ValiTools pack, and it's the one node here worth building a workflow around.

The pitch isn't the combinations themselves - plenty of nodes do those. It's that this editor tells you what happened: options get color-coded per nesting level, wildcards glow yellow when their file exists, variables highlight violet, and after a run the branches that were actually picked get white marks. You're not resolving blind; you can see the roll.

How it works

Type a prompt in the big editor widget. On run, the node resolves it in a single pass: {a|b} picks one option (with optional N:: weights, nestable), __file__ pulls a random non-empty line from a .txt in your wildcard directory (subfolders supported), and everything runs through one RNG stream seeded from the seed widget. Comments are stripped before resolution, so text inside a comment can't silently trigger a wildcard or variable assignment - an old bug, now fixed.

The clever bits are variables and the switcher. {a|b}==<name> remembers the pick; <name> then outputs the same value everywhere, even inside later combinations. ==!<name> is the silent variant - it stores without emitting, so you can tag a scene ({day|night}==!<time>) and gate whole blocks on it: <time>==day::{...} fires only when the value matches, <time>!=day::{...} for everything else. Typing < or __ opens autocomplete dropdowns (variables / wildcard files), and CTRL+Click a wildcard opens or creates its file in a built-in editor that works on remote installs.

The inputs that matter

Most of the widget list is cleanup toggles you can ignore. The ones to touch:

• seed - same seed + same prompt + same connected in1–in4 text always returns the same output. If an upstream input's text changes, the picks re-roll even with a fixed seed (it's mixed into the effective seed).
• wildcard_directory - where the .txt wildcard files live. Defaults to the pack's own wildcards folder, which ships a few example files.
• line_suffix - appends a string to every line; handy for suffixing tags with commas or periods.
• single_line_output, remove_whitespaces, remove_empty_tags - leave them on. single_line_output must be True for multi-line combinations to work.

One warning: available_loras_stem, load_loras_from_prompt, and remove_loras_pattern are compatibility placeholders kept only to preserve widget order in old workflows. This node no longer loads LoRAs - remove_loras_pattern just strips LoRA tags from the output. Want the LoRA actually applied? Add a LoraLoader.

The optional inputs are where the power is: prompt (the editor text), in1–in4 (string sockets you reference as <in1>–<in4>; they're lazy - the upstream branch only runs if the reference appears in your prompt), and vars_in (variables handed over from another VSmartPrompt, so values travel down a chain). Outputs: prompt (resolved), original_prompt, variables.

A quick example

a {photo|{oil|water} painting} of a {0.7::cute|scary} __animal__==<pet>
the <pet> naps in a (cozy:1.2) {garden|kitchen}. {day|night}==!<time>
<time>==day::{sunlight warms the <pet>|birds sing}

With the picks shown, you get something like "a photo of a cute fox. the fox naps in a cozy kitchen." - one node, one seed, a different scene every queue.

Installing

Install via ComfyUI Manager (search "ValiTools" or "comfyui-ValiTools"), or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/vangel76/comfyui-ValiTools

Then restart ComfyUI. No pip install and no model files - the pack is pure Python plus frontend JS, using only numpy/torch/PIL (already in ComfyUI) on its image-loading side.

Where people get burned

• A wildcard stays literal (__animal__ ends up in your output) when animal.txt doesn't exist in wildcard_directory. The editor highlights missing wildcards red precisely so you spot this before you queue.
• "Which prompt made this image?" is the classic dynamic-prompt pain - the workflow PNG stores the template, not the roll. Wire the prompt output into a Show Text node, and the exact resolved string gets embedded with the workflow, so you can always recover it.
• (word:1.2) weighting passes through untouched - that's ComfyUI-native, and the node doesn't interpret it. Note it's meaningless on LLM-encoded models (Flux 2, Z-Image, Anima); on CLIP-based SDXL-lineage checkpoints it still works.
• v1.5.1 fixed a real RNG bias where reseeding per pass locked different draws together. The fix means a given seed now produces a different (unbiased) result than older versions - don't expect archived seed-and-image pairs to reproduce.