Character Architect

Character Architect isn't a LoRA trainer and it isn't a character-consistency model. It's a giant, opinionated prompt-writing machine: 83 dropdowns covering appearance, ethnicity, pose, clothing families, cosplay, accessories and composition, all assembled into one coherent natural-language prompt that drops out as a plain STRING. You wire that into your text encoder instead of a text box, and suddenly "batch 40 variations of the same character skeleton" stops meaning hand-editing prose forty times.

The name "Character Architect" is doing a lot of work - but it's honest. This is the kind of thing that ships with the workflows the author actually runs, and the ones bundled in the repo are Krea 2 graphs using ClownsharkBatwing samplers. That tells you the target: LLM-encoded photoreal models that read sentences, not booru tags. If your pipeline is Illustrious or Pony, this will underperform a tag-based prompter - it emits blocks of prose, which is exactly what a Qwen or Krea-style encoder wants (see the block-structure findings in the KB's prompt-engineering notes).

How it actually works

Every dropdown offers None, Random, Forced Random, plus its concrete values. Here's the clever bit: each field marked Random resolves deterministically from the seed via a blake2b hash of (seed, field_name). Every field gets its own independent stream, so locking one property no longer reshuffles the unrelated ones around it - a real quality-of-life win over the {a|b|c} bracket wildcard trick the A1111 crowd used to live on.

The distinction between Random and Forced Random matters and trips everyone up. Ordinary Random is subject to all the hidden compatibility filters - no shoes in a close-up crop, no contradictory ethnic anatomy, one main clothing family at a time. Forced Random deliberately bypasses all of them for that one field. The frontend also adds section headers with RANDOMIZE ALL ONCE (bakes concrete values now) and SET ALL RANDOM (leaves them resolving per seed) - not the same thing, and worth keeping straight.

There's a deliberate 20% wild-card share baked into Random. Some outputs will look "off" on purpose; that's the design, not a bug.

The inputs that actually matter

Don't panic at the scroll. You set a handful, the rest can ride on Random:

• media_type, gender, content_rating - the pivot inputs, locked by default (lock_media_type, lock_gender, lock_content_rating). Note the tooltip: content rating adds a styling direction only; it never hides categories or blocks manual choices. This is an adult-content-oriented pack, handled tastefully - set it to normal if you want zero NSFW flavor.
• origin_ethnicity + ethnicity_guidance - flipping guidance on replaces the broad label with a concrete 13-profile phenotype anchor, draws eye color from an ethnicity-weighted pool, and stops random anatomy from contradicting the anchor.
• clothing_archetype - 11 style directions (Classy Chic, Streetwear, Kawaii…) that guide any row left on Random without touching your manual picks.
• seed / control_after_generate - defaults to randomize, which means every queued run re-rolls your Random rows. Set it to fixed while you tweak a single field.
• free_prompt with free_prompt_position, and the override_field + override_text pair - override is the final authority, applied after every rule, and clearing competing garment families so your connected text always shows up.

The three STRING outputs: Prompt (the full thing), Face Prompt (a face-only subset - perfect for a face-detail pass), and Inspected Value (one resolved field's final text, handy for labels, grids and stats without parsing the whole prompt).

Installation

No model downloads, no requirements.txt, no API key - it's pure Python plus a JavaScript frontend. ComfyUI Manager, search "Character Architect," or:

cd ComfyUI/custom_nodes
git clone https://github.com/InspectaGadget/ComfyUI-CharacterArchitect

Restart ComfyUI. Because there's a JS frontend that reorganizes the widgets into sections, force-refresh the browser (Ctrl+F5) after updating, or you'll see stale UI.

Where people get burned

The randomness being seed-driven is the big one: if you leave control_after_generate on randomize, your carefully locked outfit gets re-rolled every run. Lock what you care about, then fixed the seed.

Second: this is an obscure pack - there's essentially no Reddit support thread for it, so the README and the built-in tooltips are your docs. Hover any widget: every control carries its own interaction rules, probabilities and suppression logic as metadata. That's genuinely the fastest way to understand why a Random row didn't appear.

Third: the output is long by design, and LLM-encoded models have a real attention cap around 75–100 effective tokens. If a distilled model drifts with the full Prompt, trim categories or feed it the Face Prompt instead - the block structure this node produces is exactly what those encoders want, but even good structure has a budget.