WAI illustrious Character Select EN

If you generate anime on a WAI or Illustrious checkpoint, this node saves you from the single most tedious part of the job: remembering exactly how a character's tag is spelled. Reach for it when you want a known character, a specific Vocaloid, a game protagonist, a series lead, and you don't feel like memorizing 2b (nier automata), copyright suffix and all, from a list of thousands.

Why this exists

The whole reason the WAI/Illustrious family is worth using is that character knowledge is baked into the checkpoint. A character with roughly 200+ Danbooru works usually renders from its name tag alone, no LoRA required. That's the model's headline capability. The catch is that the interface is Danbooru's structured tag vocabulary, not plain English. "2B from NieR" does nothing; 2b (nier automata) is what the model actually learned. Get the spelling or the parenthetical series wrong and you get a generic girl.

This node bundles WAI's tested character list, 5,075 entries in the dropdown, so you pick from a menu instead of guessing. It's a pure text helper: it emits the correct Danbooru character tag as a prompt string. It touches no GPU, downloads no models, and does nothing on its own until you wire it into a prompt.

The inputs that matter

• character is the dropdown of 5,075 characters. Pick one, or choose random to roll one from the list.
• character_weight (0.1–2.0, default 1.0) wraps the tag in emphasis, e.g. (hatsune miku:1.2). Leave it near 1.0. Pushing toward 2.0 over-forces the character and burns the image; the useful band is roughly 0.5–1.5.
• custom_prompt (optional, multiline) is your own tags and scene. The node merges this with the character tag, and insert_before_character toggles whether your text lands before or after it.
• random_action_seed only matters when character is set to random. It decides which character you land on, so freeze it to reproduce a roll.

There are two outputs, both strings. prompt is the one you use: connect it to the positive CLIP Text Encode, or into a text-combine node if you're stitching it together with a base prompt. info is a readout. Send it to a ShowText node so you can see which character actually got picked, which is essential when you're using random.

Installing it

Easiest path is ComfyUI Manager: open the Custom Nodes Manager, search ComfyUI_Mira, click Install, and restart ComfyUI. Manually, it's a clone into your custom nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/mirabarukaso/ComfyUI_Mira

Restart ComfyUI. If it throws an error on load, run pip install -r requirements.txt inside ComfyUI/custom_nodes/ComfyUI_Mira. There are no model weights to download and no heavy dependencies; it's a text node.

Where people get burned

Wrong checkpoint. These are Danbooru tags. They mean something to WAI, Illustrious, and NoobAI, and essentially nothing to Flux, SD 1.5, or a realism model. Prompting paradigms don't cross model families. Load a WAI or Illustrious checkpoint or the character tag is just noise.

The output is only half a prompt. A character tag alone won't produce a good image. You still want a quality prefix (masterpiece, best quality) and a scene, either dropped into custom_prompt here or combined downstream. And remember it's a raw string; nothing happens until it reaches CLIP Text Encode.

Obscure characters won't resolve. Coverage is broad but not infinite. A character with only a handful of Danbooru works isn't really in the model, so the tag comes out generic and you'll need a dedicated LoRA. When a pick disappoints, that's usually the reason, not the node.

It won't fix your settings. If the character is right but the image looks off, that's the checkpoint's dials, not this node. The family wants CLIP skip 2, a sampler like Euler a, and a modest CFG in the 4–7 band. And resist the urge to pad custom_prompt: WAI's own model page warns that long quality-tag stacks and forty-word negatives actually make Illustrious output blurrier, not better. masterpiece, best quality up front and worst quality, low quality in the negative is enough. Longer is worse here, which is the opposite of old SD 1.5 habits.