Character Card

Character consistency is one of those problems people over-engineer. The practical version: you want the same character in every image this session, so the prompt-writing LLM needs to know what "Vasya" looks like - once, in one place, without you re-describing him every run. Character Card is that one place.

You fill the fields you care about and it emits a tidy Markdown block - a heading with the name, a bullet per filled attribute, free-form notes at the end. Empty fields are skipped, so you only describe what matters. That block is meant to be gathered by Context Collector and fed into an LLM's context input, and then a request that just says "Vasya and Kolya drink tea in a cafe" comes back expanded with each character's actual looks.

The inputs that matter:

• name - becomes the card heading and the name the LLM binds attributes to.
• The appearance fields: gender, age, ethnicity, eye_color, hair_color, hair_style, build, height, outfit, features - all optional, all skipped when empty.
• notes - free-form anything-else (personality, accessories, scars), as a - Notes: line.
• save_preset_as + tags - type a name and run to save the card to the pack's Card Presets library. It saves at run time, so it captures whatever's filled in whether you typed it or wired it in from outside (e.g. a photo description). Re-runs overwrite the same name.

The outputs:

• card - the Markdown block. Wire into Context Collector.
• voice - a separate typed output for the music path (only relevant if you use the pack's Siren music nodes; ignore it otherwise).

The voice output is the interesting design decision. A band member is one card: looks go to the cover-art prompt, voice_tags (a music-caption fragment like "male lead vocal, raspy baritone, close-mic") goes to Siren Cast for the song's caption, and the same card tells the lyrics LLM who it's writing for. Keeping it a separate typed output - rather than part of the Markdown - is deliberate: pouring "brown eyes, navy dress" into a music-production caption just dilutes the part that decides how the vocal sounds.

Install via ComfyUI Manager (search "Kinburg-Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Kinburg/Kinburg-Nodes

then restart. This node itself is pure string assembly - no model, no dependency - but it's built to feed the pack's Local LLM (GGUF) nodes, which do need llama-cpp-python (installed automatically by install.py) and a GGUF model of your own in ComfyUI/models/llm. The pack is a one-author personal collection with surprisingly thorough docs and a 1395-check test suite.

One trap to dodge: the card doesn't enforce anything. It's reference text, not a binding contract - the LLM still decides how faithfully it weaves the attributes in, and the diffusion model has its own ceiling on multi-character attribute binding. Fill the card, feed the context, and treat the result as "much better than nothing," which it reliably is.