Identity Mixer (Parts)

Identity Mixer (Parts) is the pack's character-factory: it builds a single reusable character description out of per-trait dropdowns - age, face type, eyes, eye color, makeup, nose, mouth, hair, skin, body type, expression, ethnicity, even nails. You assemble a person part by part, and it hands you the finished description plus a signature so you can prove two generations described the same person.

Where does this fit? The pack's generator takes an identity_profile, but that only gives you a handful of hard-coded presets. This node is the open-ended version: it reads the same trait library (identity.json, with TRAITS and curated PRESETS like Kira_Goth, Fallen_Helly, Neutral_Studio) and lets you roll your own. It's aimed squarely at the character-consistency problem our KB keeps calling the community's #1 unsolved pain - same face, hair and vibe across a dataset. No adapter, no image reference; just text discipline.

How it works

Every trait dropdown offers (random), (preset), (empty), or a specific value. (random) samples from that trait's list in identity.json; (empty) drops the trait entirely; (preset) fills all traits from a named preset. Ethnicity has a dedicated (none) that excludes it cleanly rather than leaving it blank - handy if you want the description to stay culturally neutral. The chosen parts get joined into identity_txt in a fixed order, and a SHA-1 hash of the exact combination becomes the identity_signature (first 10 hex chars). Same parts in, same signature out - that's the consistency guarantee, and it's deterministic even though (random) isn't.

The inputs that matter

preset is the shortcut: pick one and all the trait dropdowns collapse into that curated identity. Everything else (age, eyes, hair, hair_color, skin, body_type, expression_base, …) is a per-part dropdown. random_seed lets you reproduce a specific random roll. custom_intro_prefix prepends your own text (e.g. a clothing note) to the output. And if you've customized your trait library, point identities_file at your identity.json - otherwise it uses the pack's built-in one.

Outputs and wiring

Three strings come out:

• identity_txt - the comma-joined character description. This is the one you actually use.
• identity_signature - the hash, for tracking/dedup across a dataset or caption set.
• identity_meta_json - the full part-by-part breakdown as JSON.

The natural loop is identity_txt → the generator's external_identity input (or custom_intro), with lock_identity set to yes, so the flagship bakes your mixed identity into every world. For LoRA training you'd paste the same text into every caption and keep the signature in a sidecar to prove which images match which identity.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/zeeoale/PromptCreatorNode

restart, and "Identity Mixer (Parts)" appears with the rest of the pack. No model downloads and - notably for this node - no API keys at all; it's pure local text assembly.

The honest take

The trait writing is the product here, and it's good: the bundled trait lists are unusually specific ("asymmetrical side-shaved hairstyle with one side closely cropped…"), which dodges the semantic bleed you get from vague terms like "short hair". One honest caveat: text-level identity gets you consistent description, not consistent pixels - a model that can't hold a face across seeds will still drift, and for truly reliable character consistency the KB's current advice is instruction-editing (Qwen-Image-Edit) rather than prompting alone. Use this node where its text output shines: variation-with-a-locked-identity, and dataset captioning.