Wildcard Divide

This is the pack's flagship, and it's the answer to a very specific problem: you want a wildcard prompt - "roll me a random but coherent outfit" - and you want different parts of that prompt to control different parts of the image. 2girls [SEP] blonde hair [SEP] black hair is the whole pitch: the first segment applies everywhere, the second steers the left half, the third steers the right half, all in one node that also hands you the resolved text and the right canvas size. It's wildcards plus regional prompting welded together, and it's genuinely neat when it works.

How the wildcard side works

The syntax is Impact Pack's, extended. __hair__ picks a random line from your hair wildcard. A numeric prefix makes it weighted - 4, blonde, 5, black, 1, red means blonde rolls 40% of the time. Entries starting with / are patterns matched against the prompt text built so far: with /skirt/ stockings in your legs slot, stockings only enters the pool when the outfit already mentioned a skirt. You can add ! for a fallback ("choose this when the pattern does not match"), and = or =~ for exclusive or conditional-exclusive pools. It's a lot, and it's the same machinery you'd find in any decent wildcard setup - the interesting part is the region splitting.

The region side, and the inputs that matter

Write the segments into wildcard_text separated by [SEP]. Each [SEP] divides the canvas into equal parts, and you can prefix options with opt:horizontal or opt:vertical to pick the split direction, or opt:832x1216 to force a size for that roll. The first segment is the base prompt for the whole image; the rest map to their strip. That's what overall controls - on by default, meaning the base segment covers everything and the region prompts layer on top of it.

The few inputs you'll actually set:

• mode - Populate on, Fixed off. When it's on, the node re-rolls wildcard_text into populated_text every run. Flip it off to lock the current roll and stop paying for the expansion.
• seed - what makes a roll reproducible. Same seed, same outfit; bump it to reroll.
• negative - your ordinary negative conditioning, applied to the whole image.
• width / height - the canvas size, which the opt:WxH syntax can override per roll.

Outputs: positive and negative conditionings feed your sampler, model is a patched model clone you pass through instead of the original, string is the fully populated prompt (handy for seeing what actually got picked), and width/height are the resolved dimensions - wire those into your Empty Latent Image node so the canvas follows the roll. The two dropdowns, "Select to add LoRA" and "Select to add Wildcard," are conveniences that inject <lora:...> tags or __wildcard__ references into the text box for you.

Install and reality check

Same as every node in the pack: ComfyUI Manager → search "Wild Divide", or git clone https://github.com/Julian-adv/WildDivide into custom_nodes, then restart. No model downloads, no heavy dependencies. Two caveats worth knowing before you commit: the repo has been 404ing since mid-2026 (the author's GitHub now only lists an unrelated Rust project), so grab it from a mirror if Manager can't find it; and regional prompting is fiddly - the community's consensus after years of threads is that masks can bleed, each region's prompt influences the whole image rather than being airtight, and results vary a lot by checkpoint. Set expectations accordingly: it's a great way to force composition, not a reliable way to make two strangers never look at each other.