Wildcard Processor

Write one prompt template, generate a thousand different images from it. That's the whole pitch of wildcards, a trick that's been around since the A1111 days and still quietly powers most character-variation and dataset-generation workflows. Wildcard Processor is Eclipse's take on it: type A {red|blue|green} car, fix the seed, and every run picks a new combination - or the same one, if you want reproducibility.

The syntax

Three things to learn, and then you're dangerous:

• Inline options: {option1|option2|option3} picks one at random. A {young|old} {man|woman} wearing {red|blue|green} clothes - multiple sets all roll independently.
• File-based wildcards: {colors} pulls a random line from wildcards/colors.txt. This is where it scales - instead of typing thirty options, you point at a file. The pack ships an example wildcards/ folder and extracts its defaults on first launch.
• Weighted options: {red:5|blue:3|green:1} makes red five times more likely than green. Use sparingly; most prompts don't need probabilities.

Nested wildcards work too - a wildcard file whose lines reference other wildcard files - so you can build a hierarchy of reuse.

The two modes and the seed

mode is the thing that trips people up. In populate mode the node processes wildcard_text using the seed and writes the expanded result into populated_text. In fixed mode it ignores wildcard_text and uses populated_text as-is - which means you can generate an expansion, then edit the expanded text by hand and keep it. Think of it as "roll the dice" vs "keep what I rolled, then fiddle."

The seed widget has the Eclipse family's special values: -1 randomizes each run, -2 increments from the last, -3 decrements. Same rules as everywhere in ComfyUI: fix the seed and the expansion is reproducible; change it and you get new output. This is the same seed discipline as the sampler, and the same gotcha - read the number before you queue, because "randomize after generate" can silently eat your good seed.

There's also a negative_prompt input that filters comma-separated tags out of the final output after expansion - Raffle-style, so it doesn't influence selection, just scrubs unwanted tags from the result. And a wildcards dropdown for quick-inserting from your wildcard files.

Where it earns its keep

Beyond "nice variety," the two killer uses are batch character-reference libraries (one template, batch 32, get the same subject across cameras and moods with no LoRA) and dataset/prompt exploration for training. Wire processed_text into a CLIP Text Encode, connect a seed you control, and batch away. It's the kind of node that looks like a toy until you realize half your static prompts could be one template.

Install it

Part of ComfyUI_Eclipse. ComfyUI Manager → search "Eclipse" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse

On first launch the pack copies its default wildcard files into the repo's wildcards/ folder (and symlinks them under models/wildcards/smart_prompt/), and it never overwrites files you've edited. Put your own .txt files in ComfyUI_Eclipse/wildcards/. Standard deps apply; it's under Eclipse → Text.

Gotchas

The classic failures: a {filename} reference to a file that doesn't exist (the node won't silently skip it - it errors or passes the literal through, so check the path), and empty lines in wildcard files producing blank prompt segments. Also, in fixed mode the auto-update stops - if you change the seed expecting new output, nothing happens until you flip back to populate. And file-based wildcards resolve relative to the pack's wildcards/ folder, so if you've moved your ComfyUI install, re-point or recreate the junction.