Card Save

The whole point of the Kinburg pack's card system is that you describe a character or entity once and reuse it forever. The missing piece, historically, was the boring one: getting a card into the library. Before Card Save existed you'd have to route an LLM's JSON output through a JSON-extract node and hand-wire a dozen fields into a Character Card. This node closes that loop in one step.

Here's the flow the pack is built around: Grammar Presets → Local LLM (GGUF) (grammar_override) + a photo → Card Save. The grammar forces the vision model to emit a card-shaped JSON, and Card Save parses it and files it in the Card Presets library. No JSON dancing, no 12-wire choreography - the grammar's keys already mirror the card fields 1:1.

The inputs that matter:

• json_string - the card JSON. Wire an LLM output here or paste. Prose around the JSON is tolerated - the first {…} is parsed, so a chatty model's preamble won't break it.
• card_type - auto (detects character vs entity from the keys), or force character / entity if auto guesses wrong. Auto's rule: a name+description with no character attributes is an entity, otherwise a character.
• save_as - the preset name. Empty = use the JSON's own name field. Empty there too = render only, don't save. Re-saving the same name overwrites.
• tags - comma-separated labels for filtering the library later.

The outputs:

• card - the rendered Markdown block, ready to feed Context Collector in the same run.
• saved_as - the name actually used, or empty if nothing was saved.
• report - a human line saying what happened ("saved 'Vasya' (character…)", or "not saved (no name in JSON…)").

The design detail worth knowing: the node never raises. Bad or empty JSON yields an empty card and an explanatory report instead of breaking the graph - which matters because LLM output is the messiest thing you'll ever wire into a node. And the resolved name is used both as the library key and the card heading, so an explicit save_as (the real name the model couldn't read off the photo) drives the block too.

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

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

then restart. The node itself has no dependencies; the workflow it's built for needs the pack's Local LLM (GGUF) node working, which means llama-cpp-python (installed automatically by install.py, matched to your torch's CUDA version) and a vision-capable GGUF + mmproj of your own. A one-author personal pack, but maintained to a rare standard - 95 nodes, per-package docs, 1395 automated checks.