✨ windows: local tts
✨ local image gen:
→ added local server launch so the model stays in memory, no idea how stable it is, so both ways are left for now
→ test flux model support
✨ spectro analyzer for openai, google, and local tts
✨ updated gallery
✨ new music
✨ nvidia/amd switch (instead of auto detection)
✨ desktop: a tiny 'easter egg' on the main screen (5 clicks on the version)
✨ koboldcpp updated to 1.96.2
🛠 empty input field now hides after 3 sec
🛠 tweaked system contexts, so chars should be a bit less horny (just a bit)
🛠 help section updated
🛠 comic mode fixed
🛠 openrouter: fixed prompt gen for image
🗝️ changed the passwords for base, person, visionary

→ gdrive

some things about local tts:

• on first launch on windows, that annoying firewall will ask for permission. model warm-up 20-50 sec, then way faster, esp. on gpu
• drop .wav voice ref in resources/speech/, won’t work from other folders
• model restarts when changing voice
• since the model was trained on short clips, long text gets split, so you might hear pauses, but on gpu they’re barely noticeable
• didn’t add binaries to the lite windows build without local models, since they’re huge (building them was pain, probably pulls extra libs, couldn’t slim it down yet)
• imo, better quality if you use your own wav, especially for non-english, language is picked from "ai language" settings
• mac & linux planned

flux models didn’t really run for me, maybe you’ll manage. 8gb vram is not enough. not included in the main build, but if you want some pain – try it. the archive has quantized ones (might be the issue), but u can swap your own clip_l and t5xxl, game looks for them in resources\stablediff\flux, the model for flux goes in resources\stablediff\models, the game filters by name so it has to start with flux

if you’re waiting for the main story: yeah, I kinda got carried away with tech, will get back to the story stuff in one or two updates, aiming for mid-september for new content. have a chill week •ᴗ•