VERSION 2.0 BUILD 19 — the biggest release so far: 57 new features across seven areas, plus an engine-level render fix and 150+ smaller fixes from two full audits. Full breakdown in the README.
Highlights
- 🎓 LoRA Training Studio — train your own LoRA fully on this Mac: workshops with auto-captioning by a local vision model, a live loss curve, a checkpoint compare matrix against a no-LoRA baseline, publish straight into your library. Trainings queue and run unattended; a finished run can extend for more steps. Klein 4B trains on a 32 GB Mac — and the app is honest when a run wouldn't fit.
- 📚 A rebuilt, editable Library — 247 hand-tuned scenes across 19 studios, every card previewed by a real render of that exact recipe, with pinned settings that reproduce it. Hide anything, restore everything, build your own cards — or save one straight from Generate.
- 🧬 LoRA toolbox — unlimited stacking, strength sweep, bake, merge, identity cards, import by Hugging Face URL with a pre-download compatibility check,
[trigger]placeholder. - ✨ Full render control — a real 1–50 steps slider, per-tier quality (quantization) presets, seven aspect presets, up to 4 references, AI insights (ask about the image / score it vs your reference), the exact enhanced prompt revealed.
- 📦 Fast, resumable downloads — six parallel chunks, survive cancel/quit/crash, download queue, and a Models tab that stays usable meanwhile.
- 🖼 Gallery search & filters, per-LoRA render counts, batch upscales through the queue, live step + ETA on the running task.
- 🛡 Safety & honesty — a red health chip before swap/disk trouble, Repair Library, a genuinely complete "Remove all data", and a vendored engine patch that fixes renders collapsing into bands when a size transposes the previous one.
Install
- Download
Typhoonminigen.app.zipbelow, unzip. - Clear the quarantine flag once (unsigned open-source build):
xattr -d com.apple.quarantine /path/to/Typhoonminigen.app
— or double-click, let macOS refuse, then System Settings → Privacy & Security → "Open Anyway". - Open the Models tab and download a tier. Klein 4B needs no account; Klein 9B is gated (free HF token).
Requirements: Apple Silicon (M1+), macOS 14+. 16 GB RAM is comfortable for Klein 4B; 32 GB for Klein 9B and for LoRA training (4B tier).