Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

Install: force Python 3.11 venv (ship-blocker — fresh install was downloading 35GB into a 3.10 venv)

mrbizarro88515b4e5b7f
diff --git a/assets/build_phosphene_logo.py b/assets/build_phosphene_logo.py
index 55979b0..2adc8c7 100644
--- a/assets/build_phosphene_logo.py
+++ b/assets/build_phosphene_logo.py
@@ -30,7 +30,7 @@ from pathlib import Path
# Brand palette — locked colors from the prompt spec
# ---------------------------------------------------------------------------
COLORS = {
- "void": "#0A0E1B", # background
+ "void": "#00061a", # background
"yellow": "#FFE13B", # center spark
"magenta": "#FF2E9F", # inner rings (warm core)
"violet": "#B14AFF", # middle rings (primary brand color)
diff --git a/assets/logo-header.png b/assets/logo-header.png
index 497b1e8..dcd01d9 100644
Binary files a/assets/logo-header.png and b/assets/logo-header.png differ
diff --git a/assets/phosphene_favicon.svg b/assets/phosphene_favicon.svg
index dc7e236..54eeb5e 100644
--- a/assets/phosphene_favicon.svg
+++ b/assets/phosphene_favicon.svg
@@ -1,5 +1,5 @@
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 256 256" width="256" height="256">
- <rect width="256" height="256" fill="#0A0E1B"/>
+ <rect width="256" height="256" fill="#00061a"/>
<circle cx="128.0" cy="128.0" r="102.4" fill="none"
stroke="#B14AFF" stroke-width="15.36"/>
<circle cx="128.0" cy="128.0" r="25.6" fill="#FFE13B"/>
diff --git a/assets/phosphene_full.svg b/assets/phosphene_full.svg
index 4f74002..b455719 100644
--- a/assets/phosphene_full.svg
+++ b/assets/phosphene_full.svg
@@ -1,5 +1,5 @@
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1024 1024" width="1024" height="1024">
-<rect width="1024" height="1024" fill="#0A0E1B"/>
+<rect width="1024" height="1024" fill="#00061a"/>
<g data-ring="0" stroke="#FF2E9F" stroke-width="5.939" stroke-linecap="round">
<line x1="589.82" y1="512.00" x2="608.26" y2="512.00"/>
<line x1="589.45" y1="519.63" x2="607.79" y2="521.43"/>
diff --git a/assets/phosphene_mono.svg b/assets/phosphene_mono.svg
index 458079a..b69549b 100644
--- a/assets/phosphene_mono.svg
+++ b/assets/phosphene_mono.svg
@@ -1,5 +1,5 @@
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1024 1024" width="1024" height="1024">
-<rect width="1024" height="1024" fill="#0A0E1B"/>
+<rect width="1024" height="1024" fill="#00061a"/>
<g data-ring="0" stroke="#FFFFFF" stroke-width="5.939" stroke-linecap="round">
<line x1="589.82" y1="512.00" x2="608.26" y2="512.00"/>
<line x1="589.45" y1="519.63" x2="607.79" y2="521.43"/>
diff --git a/assets/phosphene_simple.svg b/assets/phosphene_simple.svg
index ac73b55..5a21415 100644
--- a/assets/phosphene_simple.svg
+++ b/assets/phosphene_simple.svg
@@ -1,5 +1,5 @@
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1024 1024" width="1024" height="1024">
-<rect width="1024" height="1024" fill="#0A0E1B"/>
+<rect width="1024" height="1024" fill="#00061a"/>
<g data-ring="0" stroke="#FF2E9F" stroke-width="12.288" stroke-linecap="round">
<line x1="591.36" y1="512.00" x2="637.44" y2="512.00"/>
<line x1="585.32" y1="542.37" x2="627.89" y2="560.00"/>
diff --git a/assets/phosphene_wordmark.svg b/assets/phosphene_wordmark.svg
index a0ae4fb..0843947 100644
--- a/assets/phosphene_wordmark.svg
+++ b/assets/phosphene_wordmark.svg
@@ -1,5 +1,5 @@
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1024 256" width="1024" height="256">
- <rect width="1024" height="256" fill="#0A0E1B"/>
+ <rect width="1024" height="256" fill="#00061a"/>
<g transform="translate(0, 0)">
<g data-ring="0" stroke="#FF2E9F" stroke-width="1.485" stroke-linecap="round">
<line x1="147.46" y1="128.00" x2="152.06" y2="128.00"/>
diff --git a/assets/phosphene_wordmark_compact.svg b/assets/phosphene_wordmark_compact.svg
index 9996c1a..5fd1121 100644
--- a/assets/phosphene_wordmark_compact.svg
+++ b/assets/phosphene_wordmark_compact.svg
@@ -1,5 +1,5 @@
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 768 256" width="768" height="256">
- <rect width="768" height="256" fill="#0A0E1B"/>
+ <rect width="768" height="256" fill="#00061a"/>
<g transform="translate(0, 0)">
<g data-ring="0" stroke="#FF2E9F" stroke-width="1.485" stroke-linecap="round">
<line x1="147.46" y1="128.00" x2="152.06" y2="128.00"/>
diff --git a/install.js b/install.js
index 292f975..7eabdc7 100644
--- a/install.js
+++ b/install.js
@@ -5,7 +5,7 @@
// pinokio.js). Every step below is safe to repeat:
//
// - clone: skipped if ltx-2-mlx/.git exists
-// - venv: Pinokio's `venv:` directive is itself idempotent
+// - venv: skipped if env/bin/python3.11 exists (force 3.11)
// - uv pip: idempotent (already-installed packages are no-ops)
// - patch: patch_ltx_codec.py is idempotent + fails loud on drift
// - hf download: resumes partial files, skips intact ones
@@ -13,6 +13,13 @@
// If the user's first install died after the venv was created but before
// the model downloads, hitting Resume Install picks up exactly where it
// left off without re-downloading the working pieces.
+//
+// IMPORTANT: we do NOT use Pinokio's `venv: "env"` directive to CREATE the
+// venv — that uses conda-base's Python (currently 3.10 on the macOS bundle)
+// which fails the ltx-core-mlx Python>=3.11 constraint. We force 3.11 with
+// `uv venv --python 3.11` and then use `--python env/bin/python` on every
+// pip step. The hf download steps still use `venv: "env"` for activation
+// only (sourcing the existing 3.11 venv to put `hf` on PATH).
module.exports = {
// Pulls in `huggingface-cli`/`hf`, `ffmpeg`, `git`, `uv`, `python3.11` etc.
@@ -41,22 +48,48 @@ module.exports = {
}
},
- // ---- Create venv + install MLX pipeline packages ----------------------
- // Pinokio's `venv: "env"` creates ./ltx-2-mlx/env/ on first run and
- // reuses it on subsequent runs. `uv pip install` is idempotent —
- // already-installed packages are no-ops. Non-editable install (no -e):
- // packages get copied into env/lib/.../site-packages/ which is where
- // patch_ltx_codec.py looks for video_vae.py.
+ // ---- Force Python 3.11 venv (SHIP-BLOCKER fix) ------------------------
+ // Pinokio's `venv: "env"` shortcut creates a venv using whatever python
+ // is on `conda activate base` — on machines where conda's base env is
+ // Python 3.10 (the current macOS bundle reality), that venv has no
+ // python3.11 and the MLX packages refuse to install:
+ // "ltx-core-mlx==0.2.0 depends on Python>=3.11"
+ //
+ // Worse, that error doesn't abort the install — Pinokio happily moves on
+ // to download 35 GB of weights into a broken venv. So we explicitly
+ // create the venv with `uv venv --python 3.11` before any pip step.
+ //
+ // Idempotency: if env/bin/python3.11 already exists we skip. If env
+ // exists but is 3.10 (our exact failure mode), nuke and rebuild — those
+ // dirs only hold a wrong-Python venv, no user data.
+ {
+ when: "{{!exists('ltx-2-mlx/env/bin/python3.11')}}",
+ method: "shell.run",
+ params: {
+ path: "ltx-2-mlx",
+ message: [
+ "rm -rf env",
+ "uv venv --python 3.11 --seed env"
+ ]
+ }
+ },
+
+ // ---- Install MLX pipeline packages into the 3.11 venv -----------------
+ // `--python env/bin/python` pins the install to the venv we just made
+ // (avoids any conda-base interference). uv pip install is idempotent —
+ // already-installed packages are no-ops on Resume Install.
+ // Non-editable install (no -e): packages get copied into
+ // env/lib/python3.11/site-packages/ which is where patch_ltx_codec.py
+ // looks for video_vae.py.
// Pin huggingface-hub>=1.0 explicitly so older Pinokio bundles still
// get the v1+ `hf` CLI used by the download steps below.
{
method: "shell.run",
params: {
- venv: "env",
path: "ltx-2-mlx",
message: [
- "uv pip install ./packages/ltx-core-mlx ./packages/ltx-pipelines-mlx",
- "uv pip install pillow numpy 'huggingface-hub>=1.0'"
+ "uv pip install --python env/bin/python ./packages/ltx-core-mlx ./packages/ltx-pipelines-mlx",
+ "uv pip install --python env/bin/python pillow numpy 'huggingface-hub>=1.0'"
]
}
},
diff --git a/launch/02_twitter.md b/launch/02_twitter.md
index 7df1378..ab86892 100644
--- a/launch/02_twitter.md
+++ b/launch/02_twitter.md
@@ -2,7 +2,19 @@
## 1. Hero Tweet
-**Variant A — technical wow (recommended)**
+**Variant C — announcement (CHOSEN — use this)**
+
+> Happy to share something I've been building — Phosphene.
+>
+> A free local video+audio generator for Apple Silicon, running LTX 2.3 in MLX. Built with a lot of help from Claude. Mostly works. There are bugs. PRs welcome.
+>
+> One-click install via Pinokio. 🧵
+>
+> [video]
+
+Tone: builder-honest announcement. Not hype, not humble-brag — straight "I made a thing, here it is, it's flawed, help me make it better." The Claude credit upfront is the differentiator: nobody else is being honest about how their AI tool was built. Lead with that.
+
+**Variant A — technical wow (alt)**
> Local generative video on a Mac just got weird.
>
@@ -12,7 +24,7 @@
>
> [video]
-**Variant B — creator vibe**
+**Variant B — creator vibe (alt)**
> made this on my Mac. no cloud, no API key, no rendering farm.
>
@@ -22,36 +34,60 @@
>
> [video]
-**Pick A.** The "joint audio+video, locally, on a Mac" claim is the thing nobody on the timeline expects to be true. Lead with the disbelief.
+## 2. Thread (reply chain to hero — paired with Variant C)
-## 2. Thread (reply chain to hero)
+**1/** What it does:
+- T2V — text → video+audio
+- I2V — image → video+audio
+- FFLF — first + last frame, model fills the middle
+- Extend — add seconds to an existing clip
-**1/** Phosphene wraps LTX 2.3 (Lightricks) running on Apple's MLX framework. The differentiator: it generates video AND audio in a single forward pass — ambient sound, footsteps, lipsync dialogue. Wan, Hunyuan, Mochi are silent. This isn't.
+LTX 2.3 generates audio jointly with video. Most local video models are silent. This isn't.
-**2/** Install is one click via Pinokio. No conda, no venv, no CUDA cosplay on a Mac.
-github.com/mrbizarro/phosphene
-
-**3/** Real numbers, M4 Mac Studio 64 GB:
-- 5s clip @ 1280×704 — ~7 min
+**2/** Real numbers on an M4 Mac Studio 64 GB:
+- 5s @ 1280×704 — ~7 min
- High quality (Q8) — ~12 min
- FFLF @ 768 — ~5 min
-Slower than an H100. Faster than a credit card.
-**4/** Modes:
-- T2V — text → video+audio
-- I2V — still image → video+audio
-- FFLF — first frame + last frame, model fills the middle
-- Extend — add seconds to a clip you already have
+Not a render farm. But: no cloud, no API key, no queue, no monthly bill.
+
+**3/** Hardware tiers auto-gate by RAM:
+- 32 GB → Compact tier (lower res, shorter clips)
+- 64 GB → comfortable floor for full 1280×704
+- 96 / 128 GB → longer clips and Q8 without pain
+
+Generative video is memory-hungry. Not negotiable.
+
+**4/** Honest disclaimer:
+Claude (Anthropic) wrote a lot of this with me. I designed, directed, reviewed, tested, broke, and fixed. There are bugs. Some edges are rough. The repo is open — file issues, send PRs.
+
+github.com/mrbizarro/phosphene
+
+**5/** Install is one click via Pinokio. No conda, no venv, no CUDA cosplay on a Mac. Search "Phosphene" in the Pinokio Discover tab.
+
+**6/** Credit where it's due:
+- LTX 2.3 model: @Lightricks
+- MLX port: @dgrauet
+- MLX framework: Apple ML team
+- Pinokio one-click runtime: @cocktailpeanut
+
+Phosphene is the panel tying them together. None of it exists without their work.
+
+**7/** Free, MIT licensed (panel — model has Lightricks' license, read it before you build a business on it).
+
+If you make something weird with it, tag me.
+If something's broken, file an issue.
+If you want to fix it, PRs welcome.
+
+Follow @AIBizarrothe for updates.
-**5/** Hardware tiers auto-gate by RAM. 32 GB runs Compact (lower res, shorter). 64 GB is the comfortable floor for full 1280×704. 96/128 GB unlocks longer clips and Q8 without pain. Not negotiable — generative video is memory-hungry.
+**8/** Roadmap: LoRAs (style, camera motion, HDR), CivitAI model browser, prompt enhance via Gemma is already shipped. Local generative video is going to get strange in a good way over the next year. Glad to be early.
-**6/** Credit where it's due. The MLX port is by @dgrauet — none of this exists without that work. The model is LTX 2.3 by @Lightricks. MLX itself is from the Apple ML team. Phosphene is just a panel on top of all three.
+**9/** If you read this far — thank you.
-**7/** It's free, MIT licensed (panel — LTX weights are under Lightricks' license, read it).
-Try it, break it, tell me what's wrong.
-Follow @AIBizarrothe for updates. RT if you ship something with it.
+Local AI tools live or die by word of mouth. A like, a repost, a quote with your take, a comment with what broke — all of it helps this reach the people who'd actually use it.
-**8/** Roadmap, briefly: LoRAs (style, camera motion, HDR), prompt enhance via Gemma is already in, CivitAI browser coming. Local generative video is going to get strange in a good way over the next year.
+RT the hero if you think a Mac dev should see it. 🙏
## 3. Reply Variants
@@ -86,7 +122,7 @@ Follow @AIBizarrothe for updates. RT if you ship something with it.
## 5. Pin the Post — Instructions
-Pin the **hero tweet** (Variant A), not the thread root individually — the pinned tweet should auto-expand the thread for new visitors. Keep it pinned for at least two weeks; local-AI Twitter discovers things on a delay, and the "joint audio+video on a Mac" hook is the strongest top-of-funnel you have. Replace it only when you ship a meaningful update (LoRAs, CivitAI integration) worth re-pinning around.
+Pin the **hero tweet** (Variant C — announcement), not the thread root individually — the pinned tweet should auto-expand the thread for new visitors. Keep it pinned for at least two weeks; local-AI Twitter discovers things on a delay. Replace it only when you ship a meaningful update (LoRAs, CivitAI integration) worth re-pinning around.
## 6. Asset Descriptions
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index e0e8fc0..08074aa 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -2112,8 +2112,12 @@ HTML = r"""<!doctype html>
<link rel="icon" type="image/png" sizes="256x256" href="/assets/favicon.png">
<style>
:root {
- --bg: #0b0e13; --bg-2: #0d1117; --panel: #161b22; --panel-2: #1c2230;
- --border: #2a3038; --border-strong: #3a424d; --text: #e6edf3; --muted: #8b949e;
+ /* Phosphene void = #00061a — the canonical dark-navy backdrop the
+ brand artwork was rendered against. Body uses it as-is; elevated
+ panels lift slightly toward blue-violet so the brand color story
+ carries from the logo through every surface. */
+ --bg: #00061a; --bg-2: #050b22; --panel: #0c1330; --panel-2: #141a3a;
+ --border: #1f2547; --border-strong: #2e3658; --text: #e6edf3; --muted: #8b949e;
--accent: #2f81f7; --accent-bright: #58a6ff; --accent-dim: rgba(47,129,247,0.18);
--success: #3fb950; --warning: #d29922; --danger: #f85149;
--radius: 10px;
@@ -2127,21 +2131,20 @@ HTML = r"""<!doctype html>
}
/* ===== HEADER =====
- Background matches the brand-void (#0A0E1B) so the logo PNG's own
- background blends cleanly into the header instead of sitting as a
- darker rectangle on a lighter panel. Logo height bumped from 40 →
- 72 so the wordmark is actually readable. */
+ Background uses var(--bg) which is the brand-void (#00061a) — same
+ color baked into the logo PNG so it blends seamlessly into the
+ header. Logo height 72px so the wordmark reads. */
header {
display: flex; align-items: center; gap: 12px; flex-wrap: wrap;
padding: 12px 18px; border-bottom: 1px solid var(--border);
- background: #0A0E1B;
+ background: var(--bg);
}
header h1 {
margin: 0; font-size: 15px; font-weight: 700; letter-spacing: -0.01em;
display: inline-flex; align-items: center; gap: 8px;
}
.brand { display: inline-flex; align-items: center; flex-shrink: 0; }
- .brand img { height: 72px; width: auto; display: block; }
+ .brand img { height: 104px; width: auto; display: block; }
.tag { color: var(--muted); font-size: 11px; }
.pill {
padding: 4px 10px; border-radius: 999px; font-size: 11px; font-weight: 500;