ꘅ Local Text

What you'd use it for

ꘅ Local Text runs a small language model locally and gives you the string. That's it. The obvious use is the one the pack is built around: write a prompt for the sampler from a rough idea, or generate a title, a caption, a variation, without touching an API key.

Why would you want a 0.6B model instead of the 8B-and-up abliterated models people usually wire into Comfy? Because it's 352 MB and runs through macOS's Core AI stack rather than holding a CUDA card hostage. In the local-LLM-node world, the standard pain is VRAM budgeting - you're sharing one card between a chat model and a diffusion model, which is why the good nodes do unload/reload dances. Here the text model lives in a helper process, the diffusion model lives in Comfy, and neither is fighting the other for 12 GB.

The honest counterweight: 0.6B is small. The author's own README says local smoke tests "exposed incorrect answers and poor decisions," and that answer-slot probabilities are uncalibrated. So this is a deterministic-ish short-rewrite tool, not a writer. If you want a model that reliably transforms a rough idea into a structured prompt, you're still looking at an 8B local model or an API.

Worth stating plainly since there's no hype to inherit: this pack has essentially zero presence in the community corpus - no threads, no discussion. It's a v0.1.0 early-alpha from an independent author (languel), and you should install it as a curiosity with a clean contract rather than as infrastructure.

Inputs and outputs

Four required inputs, no optionals:

• model - an enum with exactly one entry, qwen3-0.6b. Yes, a dropdown with one option. It's a placeholder for the pack's model catalog and you can't change it. It shares its download with ꘅ Local Decision, so if you've already pulled the model for decisions, this costs nothing extra.
• prompt - multiline STRING, defaulting to Describe a playful generative art scene in one sentence. /no_think. That /no_think suffix is a Qwen3 hybrid-thinking toggle, and leaving it in place keeps the model from burning your token budget on visible reasoning before it writes anything. Strip it only if you actually want the scratch-work.
• max_tokens - 1 to 2048, default 128. For one-sentence jobs this is generous. Raise it if you're asking for a paragraph and getting a truncated sentence.
• temperature - 0 to 2, default 0. Zero means greedy decoding: same prompt, same output, every run. That's a sane default for a prompt-engineering tool where you want to isolate one variable at a time. Nudge it to 0.7–1.0 when you want variation to pick from.

Two outputs:

• text - the generated STRING. Into your text encoder, or into ꘅ Text / JSON Viewer to just read it.
• metrics - a JSON STRING with request id, backend, model, timestamp and timings. Wire it to a second viewer while you're deciding whether 128 tokens is fast enough for your loop.

One source-level detail worth knowing: this node returns NaN from IS_CHANGED, which is the standard Comfy idiom for "always re-run, never trust the cache." So every top-level Run regenerates text - at temperature 0 you get the identical string again, which is exactly what you want when you're checking whether a workflow change broke something. The cost is that it also invalidates caching for everything downstream.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/languel/genereticore
python3 scripts/build_helper.py

Then restart Comfy, or install through ComfyUI Manager by searching GeneretiCore. Sample 02 - Local text is already wired up if you want to skip building a graph.

Two gates, and they're both hard. First, Core AI requires macOS 27+ on Apple silicon and Xcode 27 to build the helper - that's the contributor milestone, and the README is upfront that a prebuilt student release doesn't exist yet. Second, no model is bundled and nothing downloads on workflow load: click ↓ on the node to fetch the pinned Qwen3 0.6B (~352 MB, checksum verified, no Hugging Face login required). If you skip that click, you get Model missing. Download it explicitly from the node toolbar first.

If you're not on macOS 27, the same model is reachable from the command line without Comfy at all:

python3 -m genereticore.cli download --model qwen3-0.6b
python3 -m genereticore.cli text --prompt 'Describe a calm scene. /no_think' --repeat 2

That still needs the native helper, but it's a fast way to check whether the model's quality is even in the range you can use before you wire anything.

Where people get burned

ϟ versus ▷. The node's lightning button runs a local async preview and shows the answer inside the node - and updates nothing connected. So your viewer or text encoder stays untouched and the graph looks broken. Use top Run or ▷. This is the single most common confusion in the pack, common enough that the getting-started doc tabulates every toolbar button.

The download is per-model, not per-package. Text and decision share Qwen 0.6B; vision does not. Expecting one click to enable all three is how you end up staring at a missing-model error on a graph that "worked yesterday."