Phosphene 4.0 - LTX-2.5 is the engine now. Native on your Mac, with your characters and their voices.

Phosphene 4.0 is out. LTX-2.5 is now the engine everything renders on - the first native Mac build of it anywhere. Our own MLX port, our own quantized packs, mirrored ungated on GitHub: no Hugging Face login, ever.

What that means in practice:

• Fresh installs dropped to ~37 GB. LTX-2.3 became an in-panel download you only fetch if you train characters.
• Every quality tier prints a measured time from a real machine, not a guess. Balanced 5s is about 2:20 on an M4 Max; High about 4 minutes.
• LIVE PREVIEW: you watch the take forming inside the render card, and a Stop early button appears the moment the preview means something. Previews are byte-identical to renders made without them - watching costs nothing.
• Trained characters run in about 140 seconds on the Q8 pack - face AND voice - and no longer demand the High add-on. There is a face/voice split control when you want to push either on its own.
• Storyboard only loads a character's voice on shots that have dialogue. Silent shots stay clean.
• Settings gained a Storage section: what each pack costs on disk, reclaim the old generation per item, real sizes.
• Colorize, Restore, Ingredients and HDR are fixed. They were silently broken on every previous version - one step-count bug upstream of all four. A permanent gate now keeps that class of bug from coming back.
• The updater applies fixes on the FIRST click. The "click Update twice" era is over.

The clip below is a trained character running on LTX-2.5 Q8, speaking with his own trained voice.

Yesterday's post said days, not weeks. It was one day.

Weights mirror + release notes: github.com/mrbizarro/Phosphene/releases/tag/v4.0.0 - packs ship under the LTX-2.x Community License with notices riding the mirror.