Phosphene 3.8.0 is out, and it has a new tab.
STORYBOARD. Type what your film is about, pick how many shots, and a local
planner writes the whole board - every shot in the panel's own prompt
dialect, camera pinned, faces kept in frame. The shots ride the normal
render queue as drafts. You keep what works, re-roll what doesn't (the
other shots stay byte-identical), then Finish renders your keepers at
delivery quality on the drafts' own seeds. Export gives you the clips plus
a manifest.
Planning a 6-shot film takes about 25 seconds and 8 GB of RAM - and the
planner model unloads itself completely before any rendering starts, so it
never fights your renders for memory. Cast a character from your LoRAs and
every shot carries them.
FIRST: IF YOUR INSTALL IS STUCK, THIS RELEASE UNSTICKS IT. A build-tool
release (hatchling 1.32) broke every fresh Phosphene install and every
Update click at "metadata-generation-failed" - on every version. 3.8.0
pins the build correctly. If you are stuck: delete the app in Pinokio and
install fresh from Discover. Nothing else fixes it, because the broken
installer replays from disk.
Also fixed in 3.8.0:
- The in-panel update pill could stay dead forever if you installed the H3
pack (untracked files made the panel think you had local changes). Fixed- and if you were affected, the fresh install above clears it. (#52,
thanks Morac2)
- and if you were affected, the fresh install above clears it. (#52,
- Audio renders: the start-offset control now actually reaches the engine
(#46). First keyframes are cover-cropped instead of stretched (#53). - The Install Hailuo H3 menu now appears on 48 GB Macs, matching the
low-RAM engine that shipped in 3.7.0. - Honest telemetry docs: the analytics page now says exactly what is sent
(and events now tell PostHog not to derive your location). README no
longer claims "no telemetry" - that stopped being true in 3.7.0 and the
docs now match the code. Opt-out is one toggle, same as before.
Some measured numbers from the machine that built this (M4 Max 64 GB):
- Storyboard: plan a 6-shot film - 25 s, 8 GB peak, 0 resident after
- LTX Quick draft, 640x480, 5 s clip - 1:23
- LTX Balanced, 1024x576, 5 s clip - 4:57
- LTX High (Q8) with a character LoRA, 1024x576 - 5:01
- Hailuo H3 Draft, 3 s - 3:00 / High 16:9, 5 s Turbo - 8:30
AND ONE MORE THING. LTX-2.5 dropped yesterday. It is already running
natively on this Mac - MLX port, our own quantized packs, character LoRAs
carrying over, and yes, character VOICES. The clip below is bizarro
running on LTX-2.5, q8, speaking with his own trained voice. That ships
as v4.0, with the weight packs mirrored on GitHub so you never touch a
gated download. Days, not weeks.