Phosphene

Phosphene: free local AI video generator for Mac

Run LTX-2.5 and MiniMax H3 (Hailuo) on your Apple Silicon Mac. Text-to-video and image-to-video with sound, speech, lip-sync, songs and image editing. A native Mac app, offline: no cloud, no API key, no NVIDIA GPU.

Phosphene is a free, open-source AI video generator that runs locally on a Mac with M1, M2, M3, M4 or M5. It runs LTX-2.5 and MiniMax H3 (Hailuo H3) for text-to-video and image-to-video with sound in the same pass, YuE2 for songs from lyrics, and P-H3-Image, FLUX 3 Klein, Qwen-Image-Edit and FLUX.2 for images. Since 5.0 it is a native Mac app, and it still installs with one click through Pinokio. It is built on Apple's MLX: no CUDA, no cloud GPU, no API key, no credits and no watermark. Your prompts, pictures and videos never leave your Mac.

Current release: v5.0.8. Made by Mr Bizarro.

Why Phosphene

• Free, with no credits. No subscription, no sign-up, no watermark, no per-render charge. Make as many takes as you like.
• Private and offline. Rendering happens on your Mac. Download the models once, then work without a connection.
• Sound comes with the picture. Footsteps, rooms, voices and lip movement are generated together, not added later.
• One app for the whole film. Shots, long takes, speech, songs, stills, trained characters and an editing timeline.
• Made for the Mac. Phosphene checks your memory and only offers what your Mac can run.

What you can make

Text-to-video and image-to-video with sound (LTX-2.5 on Apple Silicon)

Turn a prompt or a picture into a video clip with sound, using the MLX port of Lightricks' LTX-2.5. Set first and last frames, guide a shot with 3–8 keyframes, Extend a clip, Retake one moment, or lip-sync a face to your own voice or song. Upscale & Face Fix is built in. Works from 16 GB of memory.

Run MiniMax H3 (Hailuo) locally on a Mac

MiniMax H3, also called Hailuo H3, makes scenes where people talk, with lips that match their voices. Use keyframes, extend a shot, lip-sync, or Add sound to give a silent clip its own soundscape. Pika H3 Flex, Pika Labs' six-step H3, is a Speed choice: about 2.2× faster than Best at the same size, a preview. About 75 GB, one click from the Pinokio sidebar, from 16 GB of memory. Its licence excludes the EU, UK, South Korea and US.

Long AI films: One Shot, Storyboard and the Editor

• One Shot: one continuous shot of 30 seconds to 2 minutes that never cuts.
• Storyboard: turn one idea into a multi-shot film. Plan the shots, render them, keep the best takes, or turn on Auto. Add a song to cut a music video to the beat.
• Editor: trim, split, J-cuts and L-cuts, music, titles and transitions. Render one file or send the timeline to DaVinci Resolve, Premiere Pro or After Effects.

Remix your own footage

• Motion Control: copy the motion, camera move and pose of a clip you give it while your prompt paints a new subject and scene (guide).
• Ingredients: compose 2–8 reference pictures (a character, a prop, a place) into one clip.
• Colorize: turn black-and-white footage into colour.

Speech and lip-sync

Turn text into a spoken voice, then lip-sync a face to it, all on your Mac.

Songs from lyrics with YuE2: a local Suno alternative

Music Studio uses YuE2 to write a full song with vocals and instruments from your lyrics and a style, or an instrumental. Make another take, cover a song, or get the score. About 11 GB, installed inside the app, from 16 GB of memory.

AI images with P-H3-Image, FLUX 3 Klein, Qwen-Image-Edit and FLUX.2

New in 5.0: P-H3-Image by Pruna AI makes 8-megapixel pictures with exact text (needs MiniMax H3 and 48 GB or more). FLUX 3 Klein by Black Forest Labs makes an image from a sentence in about 50 seconds on an M4 Max, or edits the reference pictures you give it. Qwen-Image-Edit and FLUX.2 (via mflux) make images and edit them in words, like “make the jacket red”, with up to three reference pictures. Ideogram 4 puts readable words on a picture. Send any still straight to video.

Train a character LoRA on your Mac (face and voice)

Train an LTX-2 character LoRA of an original character, or of yourself, on your Mac: face from your photos, voice from your recordings. Local Gemma writes the training captions for you. Browse and install LTX LoRAs from CivitAI inside the app.

Scripts and AI agents

Add Phosphene to Claude Code, Codex or any MCP client in one line (Phosphene running in Pinokio):

claude mcp add phosphene -- ~/pinokio/api/phosphene.git/ltx-2-mlx/env/bin/python ~/pinokio/api/phosphene.git/agents/phosphene_mcp.py

There is also a phosphene command (agents/phosphene_cli.py); both are in docs/AGENT_TOOLS.md. Under them, everything the app does is reachable over a local HTTP API (docs/API.md), so your own script can queue renders too. There is also an image layout API for agents (docs/AGENT_API.md). Gemma runs locally for prompt Enhance, captions and the storyboard planner.

Mac requirements: M1 to M5, memory and disk

Any Apple Silicon Mac, M1 or newer, including MacBook Air, MacBook Pro, Mac mini and Mac Studio. Memory decides what you can run. Intel Macs and Windows are not supported.

Memory	What you can make
16–24 GB	LTX-2.5 video from text or a picture at smaller sizes, images, and songs. MiniMax H3 from 16 GB (Draft and Standard clips; High from 18 GB). Close other apps and browser tabs before you compose on a 16 GB Mac.
32–36 GB	Everything above.
48–64 GB	LTX keyframes, Extend, trained characters with Q8 weights, and the High tier.
96 GB+	Most modes at full size (from 80 GB). Size limits come off at 120 GB.

Disk: setup downloads no models; allow 18 GB free, including 15 GB kept free. Pick your models inside the app: LTX-2.5 Q4 video is about 27.5 GB, image models download on first use. MiniMax H3 (about 75 GB) and Music Studio (about 11 GB) are optional.

How fast is it on a Mac?

On an M4 Max with 64 GB, a 5-second LTX-2.5 clip at High takes about 6 minutes, a 5-second MiniMax H3 clip on Draft + Fast about 5 minutes, and a 3-minute song about 3 minutes. Smaller Macs take longer: across Phosphene users, a 5-second LTX-2.5 clip at Balanced (480p) took a median 10m40s on a 16 GB M4. Phosphene estimates the time before you press Generate. Bench measures your own Mac and compares it with other Macs; see the benchmarks.

Install: the Mac app, or one click with Pinokio

Mac app (recommended): download Phosphene for Mac, open it and pick your models. It can reuse models your Pinokio install already downloaded.

Pinokio: Phosphene also installs with one click through Pinokio (pinokio.computer), a free app launcher:

1. Install Pinokio.
2. Open Discover → Download from URL and paste https://github.com/mrbizarro/phosphene.
3. Press Install.
4. Press Start, then open the app.

To update, press Update in Pinokio, then restart Phosphene. The docs have the step-by-step install, troubleshooting and every feature in detail. To install by hand instead, follow the manual install. Linux on Apple Silicon (docs/LINUX.md) is maintained by @joshuaswarren on a best-effort basis.

Pro

Phosphene is open source and stays free. Pro is optional, for people who use it every day: Faster Flex Turbo for quicker renders, new features first, and priority support. Everything available today stays free, Phosphene stays open source under MIT, and new model ports reach everyone free. Prices and what is on sale now: phosphene.video/pricing.

FAQ

Is Phosphene really free? Yes. No subscription, credits or watermark, and every feature available today stays free. Pro adds Faster Flex Turbo and early access; it does not lock today's features.

Does it work offline? Yes. Download the models once, then render without a connection. When online, it checks GitHub for updates and downloads models you ask for.

Can I run LTX-2 or MiniMax H3 on an M1, M2 or 16 GB Mac? LTX-2.5 runs from 16 GB at smaller sizes on any Apple Silicon Mac. MiniMax H3 runs from 16 GB (Draft and Standard; High from 18 GB). The Mac requirements table above has the details.

LTX-2.5 or MiniMax H3: which should I use? MiniMax H3 for people talking: dialogue and voices that match lips. LTX-2.5 for everything else: more modes, keyframes, trained characters, Motion Control and smaller Macs. Phosphene has both in one switcher.

Are LTX-2.5 and MiniMax H3 open source? Their weights are open to download, each under its own community licence (see License below). Phosphene's code is MIT.

Does it run on Windows or Intel Macs? No. Phosphene is built for Apple Silicon. On Windows with an NVIDIA GPU, look at ComfyUI or WanGP.

Can I run Wan 2.2 or FramePack on a Mac? Phosphene doesn't include them. Draw Things lists Wan 2.2, and mlx-video runs Wan 2.1 and 2.2 from Python. Phosphene's LTX-2.5 and MiniMax H3 cover the same jobs with sound.

What is the best local AI video generator for Mac? How is Phosphene different from ComfyUI, Draw Things or LTX Desktop? ComfyUI gives you a node graph for every model; Phosphene gives you ready-made workflows built on MLX for the Mac. Draw Things has an App Store app and cloud compute; LTX Desktop comes from the makers of LTX. Each is ahead of Phosphene somewhere; pick the one that fits how you work.

Is it an alternative to Veo 3, Runway, Kling or Higgsfield? Those run in the cloud and charge per render. Phosphene runs open models on your own Mac for free. Cloud models are faster, and some are sharper; Phosphene is private, unlimited and yours.

Can I use what I make commercially? Phosphene's code is MIT. Each model has its own licence, and some are non-commercial (see License below). Check a model's licence before you sell its output.

Privacy

No prompt, picture, video or filename is uploaded. The app listens on 127.0.0.1 only, with no login, so don't expose it to a network. Phosphene checks GitHub for updates and contacts Hugging Face or CivitAI only when you download models or LoRAs. Anonymous usage counts (version, hardware class such as “M4 Max / 64 GB”, and render stats) are on by default; one click in Settings turns them off, and every field is listed in docs/ANALYTICS.md. Make original or stylised characters, or yourself with documented consent; follow the acceptable use policy on phosphene.video.

Contributing

Bug reports, ideas in Discussions and pull requests are welcome. A report with your chip, macOS version, Phosphene version and the error line is worth as much as a patch. Contributors are credited in the release notes. See CONTRIBUTING.md.

Credits

• Lightricks for LTX-Video 2.x, and @dgrauet for ltx-2-mlx, the MLX port Phosphene builds on.
• MiniMax for Hailuo H3, and Apple ML research for MLX.
• LightX2V for lightx2v v1.3, the adapter for Best on High; TaoLive (TaoMate-H3) and @kijai for the 3-step Fast adapter.
• Pika Labs (Mellis, Inc.) for Pika H3 Flex, a MiniMax H3 Model Derivative, downloaded from Pika's own Hugging Face repo at a pinned revision and converted on your Mac. Powered by MiniMax H3.
• Black Forest Labs for FLUX 3 Klein, and mflux-community for its MLX conversions.
• Pruna AI for P-H3-Image.
• @filipstrand for mflux; Multimodal Art Projection for YuE2 and @vanch007 for its MLX port.
• mlx-community for Gemma on MLX, and @cocktailpeanut for Pinokio.

Thanks to @blackest, @tanis2000, @joshuaswarren, @sahilkashyap64, @Morac2, @ybekocak, @PhantombrainM, @Viktorminator, @Townesvz, @Sebartus, @mrapolo2020, @Techy2023, and everyone who filed a bug.

License

Phosphene's own code is MIT. The AI models are not part of this repo and each has its own licence. Phosphene shows a model's licence before download or first run when the licence requires it.

Ideogram 4, YuE2's weights, FLUX.1 Krea [dev] and Bria FIBO are non-commercial. YuE2 lets individual creators sell songs they make. Hailuo H3's licence excludes the EU, UK, South Korea and US. Check a model's licence before using its output commercially. The licence texts are in the LICENSES folder.