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.

Unpin ltx-2-mlx — audio bug was MLX 0.31.2, not dgrauet's commits

mrbizarroe45ff649a4cc
diff --git a/install.js b/install.js
index d7ed5a8..c971afa 100644
--- a/install.js
+++ b/install.js
@@ -48,29 +48,16 @@ module.exports = {
}
},
- // ---- Pin ltx-2-mlx to dcd639e (SHIP-BLOCKER fix) ----------------------
- // SHIP-BLOCKER history: dgrauet/ltx-2-mlx upstream made 100+ commits
- // between the 0.1.0 baseline and the 0.2.0 release HEAD. Several
- // touched the audio path (570cce8, ca784dd, 22adef1) and empirically
- // produced -22 dB output amplitude — peaks at -37 dB vs the working
- // baseline's -15 dB peaks. The "fixes" claim to align MLX numerics with
- // PyTorch reference, but the LTX 2.3 weights were trained against the
- // pre-fix MLX behavior, so post-fix output is quieter / hissy.
- //
- // dcd639e is the original "complete MLX port" commit at version 0.1.0,
- // which the user verified produces the expected audio levels. We pin
- // the working tree there explicitly. Bisecting the 100-commit range
- // for the actual regression is post-launch work.
- //
- // Idempotency: `git checkout` is safe to re-run on every install /
- // resume install — already-on-dcd639e is a no-op.
- {
- method: "shell.run",
- params: {
- path: "ltx-2-mlx",
- message: ["git fetch origin", "git checkout dcd639e"]
- }
- },
+ // ---- ltx-2-mlx version: use HEAD (the original audio bug was MLX 0.31.2,
+ // not anything dgrauet shipped). Earlier we briefly pinned to
+ // dcd639e thinking the audio regression was in dgrauet's commits;
+ // empirical follow-up showed pinning mlx==0.31.1 alone fixes audio
+ // on HEAD. dcd639e was missing several APIs the panel calls
+ // (cfg_scale on extend_from_video, the 0.2.0 I2V structure our OOM
+ // patch targets, split_model.json filename resolution). HEAD with
+ // the mlx pin gives us working audio AND working Extend / I2V.
+ // Leaving this comment as a marker so we don't re-introduce the
+ // pin without re-verifying the audio path. ----
// ---- Force Python 3.11 venv (SHIP-BLOCKER fix) ------------------------
// Pinokio's `venv: "env"` shortcut creates a venv using whatever python
@@ -159,22 +146,9 @@ module.exports = {
}
},
- // ---- Symlink transformer.safetensors -> transformer-distilled.safetensors
- // SHIP-BLOCKER history: at the dcd639e pin (0.1.0), the loader hardcodes
- // load_split_safetensors(model_dir / "transformer.safetensors", ...)
- // But the published Q4 model repo uses the 0.2.0+ naming convention with
- // split variants (transformer-distilled.safetensors, transformer-dev.
- // safetensors, transformer-distilled-1.1.safetensors). The 0.2.0+ code
- // resolves the right variant from split_model.json; the 0.1.0 code does
- // not. A relative symlink bridges the gap with zero disk overhead.
- // `ln -sf` is idempotent (overwrites stale symlink, no-op if correct).
- {
- method: "shell.run",
- params: {
- path: "mlx_models/ltx-2.3-mlx-q4",
- message: ["ln -sf transformer-distilled.safetensors transformer.safetensors"]
- }
- },
+ // ---- (no transformer.safetensors symlink needed on HEAD — 0.2.0 reads
+ // split_model.json to resolve transformer-distilled.safetensors.
+ // Symlink was a workaround for the dcd639e pin we no longer use.) ----
// ---- Download Gemma 4-bit text encoder (~6 GB) ------------------------
{
diff --git a/update.js b/update.js
index cff5b07..dff8829 100644
--- a/update.js
+++ b/update.js
@@ -5,16 +5,16 @@ module.exports = {
method: "shell.run",
params: { message: "git pull" }
},
- // Re-pin ltx-2-mlx to dcd639e — DO NOT `git pull` blindly here.
- // upstream's 0.2.0 regressed audio amplitude by 22 dB; until we bisect
- // the actual regression and either patch it or move forward to a
- // post-fix commit, we hold dcd639e steady. install.js does the same.
- // git fetch is safe (just updates remote refs); checkout is idempotent.
+ // Pull ltx-2-mlx HEAD. (We previously pinned to dcd639e thinking audio
+ // regressed in dgrauet's commits; turned out the audio bug was in mlx
+ // 0.31.2 itself. HEAD has the APIs the panel needs — cfg_scale on
+ // extend_from_video, the I2V structure our OOM patch targets, etc.
+ // git fetch + checkout main is idempotent.)
{
method: "shell.run",
params: {
path: "ltx-2-mlx",
- message: ["git fetch origin", "git checkout dcd639e"]
+ message: ["git fetch origin", "git checkout main", "git pull --ff-only origin main"]
}
},
// Force-downgrade mlx to 0.31.1 — fixes 22 dB audio regression on mlx