Argus docs
Argus is a suite of small, separately released tools that take images from raw to a trained and evaluated LoRA, and then put it to work. Each stage is its own Python package with its own HTTP service, and Argus Studio is the web UI that ties them together:
- argus-quarry acquires public-domain and CC0 images with their provenance (Gallery).
- argus-curator scores, de-duplicates and moderates them into a training set (Curate).
- argus-lens writes structured, training-ready captions (Caption).
- argus-forge turns a curated export into a ready-to-run training config (Forge).
- argus-proof evaluates the trained LoRA and feeds findings back (Proof).
- argus-atlas bakes a georeferenced fantasy map from real terrain (Atlas).
- argus-darkroom generates images with a guard before and after, and a record of every attempt (Darkroom).
- argus-cortex is the shared foundation the others build on: the taxonomy, wire-schema versioning, moderation clients and the optional lineage, vector and blob stores.
These pages cover the whole suite. The concept pages for the Caption tool's controls stay next to the tool, in Studio: target style, target backend, target category and prose enrichment.
Setup and services
The packages lean on services they don't ship: ComfyUI, Ollama, NudeNet, SingGuard, Hugging Face models and a GPU host.
- Setup and services: one page per service, with who uses it, the env vars, the minimal setup and the official guides.
- Moderation: how the suite judges images and prompts, across curator, darkroom, cortex and proof.
- What to run for which UI: what each demo UI needs live, and the no-GPU command.
Packages
Each package's own README and docs/ pages, imported from its dev branch at
a recorded commit. Fixes go to the package, never to the copy here.
A file that fails the docs opsec check upstream is held back, and its package page links it on GitHub instead.
Demos
Every demo replays recorded output, so none needs a GPU or a backend: