Free AI video with sound on Apple Silicon. Download the models once, then make films offline without uploading a thing.
A dark stage and a lit one. Phosphene joined the stills with keyframes, then extended the shot.
The camel bells weren't added in post.
Write a shot, or let H3 fill the gap between two frames.
Give a still something to do.
Make an image in Phosphene, then give it a part. Even a bust can sing.
The soundtrack gets a face.
Music Studio made the song. H3 lip-synced Bizarro to it on the same Mac.
Every frame stayed in the room.
One Mac rendered the picture and the sound. Nothing on this timeline was filmed or uploaded.
LAST TAKE: 64 seconds, made entirely in Phosphene.
Pick a shot from the film and play it with sound. All 11 clips were made in Phosphene on the same Mac.
01 · Hailuo H3 · Image to video + Extend x2 · 7.0 s
A still becomes a 15-second push-in, with two extensions.
02 · Hailuo H3 · Keyframes + Extend · 5.0 s
H3 joined the dark and lit stills with keyframes, then extended the shot.
03 · Hailuo H3 · Text to video, dialogue · 8.5 s
Picture and voice, complete with wind and camel bells. One prompt did the lot.
04 · LTX-2.5 · Text to video · 5.5 s
A paragraph sinks the ballroom. You can hear the diver breathe inside the helmet.
05 · Hailuo H3 · Keyframes · 6.0 s
Pencil sketch in, storm out. The render fills the gap between the two frames.
06 · Hailuo H3 · Image to video · 5.0 s
One still image, and the lanterns start to rise.
07 · LTX-2.5 · Lip-sync · 5.5 s
A bust gets a singing part. One photo, lip-synced to a song.
08 · Music Studio + H3 · Song + lip-sync · 8.0 s
Music Studio made the song. H3 lip-synced Bizarro to it.
09 · LTX-2.5 · Text to video · 3.5 s
Bizarro between renders, with a trained face and a trained voice.
10 · Hailuo H3 · Fast HD + Upscale & Face Fix · 5.0 s
Rendered in H3 Fast HD; Upscale & Face Fix handled the finish.
11 · Hailuo H3 · Lip-sync · 10.0 s
The final cue: "One last take, and then I'm gone."
Light patterns you see with your eyes closed.
57,000+
renders completed on Macs since August
52,000+ video renders 3,500+ image jobs 700+ song renders
0 estimated renders since you opened this page, based on this week’s rate. Not live data.
Updated Oct 20, 2026. Phosphene’s anonymous usage stats can be switched off, so actual render totals are higher. Pinokio counts installs on its project page. Every number is rounded down.
Your Mac's memory decides which doors open.
Choose your memory to see what you can make. Find it under Apple menu → About This Mac → Memory. Phosphene checks this for you when it starts.
Available at this memory: 5 of 10
Apple Silicon, M1 or newer, from 16 GB. Set aside about 40 GB of disk for the first install, plus 75 GB if you add Hailuo H3. Intel Macs and Windows aren't supported. Read the memory guide.
Every feature available today stays free. Supporter and Pro are coming for people who want to back the project or get new tools. Neither is on sale yet.
Free today. Still free tomorrow. No feature gets taken away and sold back to you.
$0 forever · MIT
$5/mo or pay what you want from $3 · $50/yr
$12/mo or $99/yr
At launch: 200 Founding Pro lifetime seats, $249 each.
Join the waitlistSupporter and Pro are coming soon. Prices and benefits above are planned; neither is on sale. Pro features haven't shipped. Everything in Free today stays free.
For now, the waitlist buttons link to @PhospheneAI on X for signup news. Following on X doesn't register you for a waitlist.
Tell Claude Code, Codex or another AI agent what you want to make. Your agent sends the job to Phosphene and follows the render. The picture and sound are made right on your Mac.
curl -s 127.0.0.1:8198/status{"running": false, "current": null, "queue": [], …}curl -s -X POST 127.0.0.1:8198/queue/add \
-d mode=t2v -d engine=h3 -d h3_quality=high -d h3_length=10s -d seed=-1 \
--data-urlencode "prompt=Live-action, cinematic, golden hour in the open desert. A medium close-up of a weathered old caravan leader with a…" \
--data-urlencode 'h3_chain_prompts=["He looks straight into the lens and (S1), in a deep, gravelly,…", "He holds the gaze, a faint tired smile in his…"]'{"ok": true, "id": "j-7f3c…"}curl -s 127.0.0.1:8198/status # Checking progress"phase_label": "Denoising · step 5 / 10", "pct": 41{"status": "done", "elapsed_sec": 1107, "output_path": "…/mlx_outputs/the_caravan.mp4"}POST /oneshotMake a whole film, 30 to 120 seconds long, from one request.
GET /oneshot/optionsCheck which engines, film lengths and quality settings your Mac supports, plus its trained characters.
POST /queue/add · GET /statusRender one shot, follow its progress and get the finished file's location.
POST /image/agentMake an image with text and objects placed where you want them.
Any agent that can make HTTP requests can use Phosphene. The repo includes AGENTS.md, CLAUDE.md, GEMINI.md and QWEN.md with instructions.
The code is open. Pull up a chair.
Mr Bizarro made Phosphene. People trying things on their own Macs keep making it better. The app is open source under MIT. Read the code, fix a bug, or take it somewhere unexpected.
with thanks to
built on
directed by
Yes. Your Mac makes the video and sound, songs and images. Download the models once, then you can render offline. Your prompts, photos and finished work stay on your machine. When online, the app checks GitHub for updates every 30 minutes, downloads models you request and sends anonymous usage counts. You can turn those counts off in Settings.
Start with an Apple Silicon Mac, M1 or newer, and at least 16 GB of memory. That runs LTX-2.5 at smaller sizes. Songs need 24 GB; Hailuo H3 starts at 36 GB. At 48–64 GB, more modes and character training options open up. Phosphene checks your memory and shows what fits. Intel Macs and Windows are not supported.
Set aside about 40 GB for the first install, which downloads roughly 37 GB of models. Add about 75 GB if you want Hailuo H3 and 11 GB for Music Studio. Downloads resume if your connection drops.
On an M4 Max with 64 GB, a 5-second LTX-2.5 clip at High takes about 4–5 minutes. A 5-second Hailuo H3 clip on Draft + Fast takes about 5 minutes; a 3-minute song takes about 3. Your settings and Mac matter. Phosphene estimates the time before you press Generate and learns from your renders.
Actually free. No subscription, credits or charge for another take. The app is open source under MIT, and every feature available today stays free. The AI models have their own licences; check those before using their output commercially. Supporter and Pro are planned, but neither is on sale.
For making short AI video clips with dialogue and sound on your own Mac, yes. Phosphene runs LTX-2.5 and Hailuo H3 locally, with no cloud account or credits. Expect different results and render times from a cloud service. The gallery shows what these models can do.
Yes. Open Train Character and start with 15 to 500 photos. Gemma captions them locally, and you can train a voice in the same run. The full training recipe needs 64 GB. Smaller Macs use a lighter recipe, and Phosphene checks how strongly the training changed the model.
No. Pro is for new additions only: speed-ups and optimizations, early access to new model ports, optional cloud models with your own keys, editing exports, batch jobs and scheduling, multi-character screenplay-to-film, licensed character packs and automation tools. Speed-ups, optimizations and big new work, including model ports and heavy engine work, go to Pro first for at least 60 days. Whether or when they reach Free is decided case by case. Planned pricing is $12/month or $99/year, with 200 Founding lifetime seats at $249 at launch. Nothing is on sale yet.
Supporter is coming for exactly that. The plan is $5/month, pay what you want from $3, or $50/year. You get a badge, betas a week early, a roadmap vote and an optional credit. It does not include Pro features and is not on sale yet. The pricing page has waitlist details.
There is a community-maintained port for Linux on Apple Silicon M1/M2, thanks to @joshuaswarren. Support is best effort. The Pinokio installer is for macOS.
Now make the film in your head.