A model on its own is half an answer. These group the checkpoint, the data it was trained on, the benchmark it is measured against and the demo that shows it working.
Everything needed to tag an illustration corpus: the tagger, the vocabulary it was trained on, the benchmark it is measured against, and a demo to sanity-check all three.
Updated 28 Aug 2026
The PrismStream line: the checkpoint, the paired line-art data behind its conditioning, and the demo where you can paint a region and argue with it.
Updated 20 Aug 2026
What we use to decide whether a change was real: a preference-trained scorer, human preference pairs, a stratified benchmark, and the ranker for eyeballing the result.
Updated 11 Aug 2026
Small enough for CPU inference and honest about it. Every entry here has been measured on a four-year-old machine rather than an A100.
Updated 30 Aug 2026