CLIPTextEncodeWildcards3

This is the "kitchen sink" version of the pack's wildcard encoder: one node, one seed, both sides of the conditioning at once. You type a wildcard-enabled positive prompt into positive, a wildcard-enabled negative into negative, plug in clip and a seed, and you get two CONDITIONING outputs - one for each - ready to drop into a KSampler with no intermediate nodes between.

Where it earns its keep is batch work where you want the positive and the negative to stay in sync. Say your wildcard files hold styles and your negative holds the "this style's worst failure" counterpart. Because a single seed drives both expansions, one draw gets you a coherent pair - a style and its matching negative - instead of a random positive bolted to an unrelated negative. That coupling is the whole reason to use the 3 variant over two separate wildcard encoders.

Inputs and outputs

Four inputs:

• clip - from your checkpoint or CLIP loader.
• positive (multiline) - the wildcard prompt for what you want.
• negative (multiline) - the wildcard prompt for what you don't.
• seed (INT) - pins both random draws; the node seeds Python's random module once, so both expansions come from the same draw.

Two outputs, both CONDITIONING: the first is the encoded positive, the second the encoded negative. Wire them to the sampler's corresponding slots.

Mechanically it's the same engine as the pack's other encoders - wildcards.run() expands each field, clip.encode() turns it into conditioning. Card files live in the wildcards folder at your ComfyUI root, one level above custom_nodes, and the usual syntax applies: __name__ pulls a random line from name.txt, {a|b|c} picks an option, {2$$a|b|c} picks two.

Install and caveats

Clone, restart, done. ComfyUI Manager can install it if you search "Lilly":

cd ComfyUI/custom_nodes
git clone https://github.com/lilly1987/ComfyUI_node_Lilly

No models to fetch; rich and chardet get pip-installed on first load. Two notes grounded in how the pack actually behaves: changing the seed is what re-rolls the draw, because ComfyUI caches unchanged inputs, and a negative prompt only does its job if your model actually uses guidance - at CFG 1 on distilled models the negative box is dead weight regardless of which node feeds it. Also, this pack is unmaintained; the author recommends Impact Pack's wildcard node for new work. Fine to lean on this one inside an old workflow that already imports it.