Phosphene
A local AI video studio runs Hailuo H3 and LTX-Video on Apple Silicon Macs, with joint audio generation and character LoRA training.
Phosphene is a generative media studio for people creating videos on their own Mac. It combines video, still images, music and character training in one interface, with generation running on-device.
- Video with sound: Generate clips from text, images or audio using LTX-Video 2.5 or MiniMax Hailuo H3, including dialogue, lip-sync and ambient sound.
- Character training: Train LTX character LoRAs for face and voice from one dataset; Hailuo H3 supports imported LoRAs but doesn't offer in-panel training.
- Shot control and remixing: Guide clips with first and last frames, extend existing shots, reuse motion from a reference video, combine reference images or colorize footage.
- Still images: Create and edit images with FLUX.2 and Qwen-Image-Edit.
- Music generation: The optional local YuE2 engine creates vocals and arrangements from lyrics and a style description, and its songs can feed the video workflow.
Phosphene runs on macOS with Apple Silicon through MLX, without CUDA, cloud generation or API keys. Prompts, images, videos and filenames stay on the Mac. It sends anonymous usage counts and hardware and render statistics unless you disable analytics.
Available workflows depend on unified memory. Advanced LTX modes require at least 48 GB; Hailuo H3 requires at least 36 GB with its compact engine, and YuE2 requires 24 GB.
The app's code is open source under MIT. Model weights have separate terms: LTX uses the LTX-2.x Community License, and Hailuo H3 uses the MiniMax Community License with territory restrictions. YuE2 weights use CC BY-NC 4.0 with permission for individual creators to monetize songs; companies need a weights license.