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.

Y1.036 route Extend to Q8 dev transformer + recalibrate Comfortable times

mrbizarro78241dc0fcc0
diff --git a/README.md b/README.md
index 9eb2770..daa3400 100644
--- a/README.md
+++ b/README.md
@@ -97,7 +97,7 @@ and step count change.
|---|---|---|---|
| **Quick** | Q4 distilled | 640×480 · ~2 min | Fast sanity checks. 4:3 native output. |
| **Balanced** | Q4 distilled + Lanczos export | 1024×576 → 1280×720 · ~4–5 min | Default. Standard 16:9 delivery without cropping; good first bet for faster HD-looking clips. |
-| **Standard** | Q4 distilled | 1280×704 · ~7 min | The daily driver. Q4 weights (~20 GB on disk). |
+| **Standard** | Q4 distilled | 1280×704 · ~8 min | The daily driver. Q4 weights (~20 GB on disk). Boost/Turbo Speed shaves ~23 % / ~34 %. |
| **High** | Q8 two-stage + TeaCache | 1280×704 · ~12 min | Sharper detail, fewer artifacts on faces and text. Optional Q8 download (~37 GB extra). Required for FFLF. |
The **Export** control can also run a lightweight ffmpeg Lanczos pass after
diff --git a/VERSION b/VERSION
index fa474b3..37955f2 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-Y1.035
+Y1.036
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index 00c0dfb..d45f484 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -1845,18 +1845,31 @@ CAPABILITIES: dict[str, dict] = {
"of memory above that. 768 is a sweet-spot working size; "
"you can run a separate upscaler afterwards if you need 720p+."
),
+ # Times are measured wall clocks for a 5 s render (121 frames @ 24 fps)
+ # at Exact speed, on the canonical Comfortable hardware (M-Studio /
+ # M-Max 64 GB). Y1.034 + Y1.035 added a temporal-streaming VAE decode
+ # patch that adds ~30 s of decode work on a 5 s clip in exchange for
+ # not melting on long ones; that bump is reflected in `standard` going
+ # from ~7 to ~8 min vs. pre-Y1.034 baselines (median 459 s observed
+ # across 26 Y1.013-Y1.024 runs; Y1.035 observed 493 s).
+ # Boost/Turbo Speed pills shave ~23%/~34% off Standard respectively.
"times": {
"t2v_draft": "about 2 min",
- "t2v_standard": "about 7 min",
- "i2v_standard": "about 7 min",
+ "t2v_standard": "about 8 min",
+ "i2v_standard": "about 8 min",
"high": "about 12 min",
- "keyframe": "about 5 min (at 768 px)",
- "extend": "about 11 min (at 768 px)",
+ "keyframe": "about 6 min (at 768 px)",
+ # Extend on Comfortable measured 16 min for +3 s at 768 px on
+ # M-Max 64 GB (Q8 dev transformer, CFG=1.0, 12 steps). The
+ # earlier "about 11 min" estimate predates the Y1.036 Q8 routing
+ # — pre-Y1.024 Extend ran on Q4-distilled-by-accident (faster
+ # weights, but technically loading the wrong model).
+ "extend": "about 16 min (at 768 px, +3 s)",
},
"quality_times": {
"quick": "~2 min",
"balanced": "~5 min",
- "standard": "~7 min",
+ "standard": "~8 min",
"high": "~12 min",
},
},
@@ -2629,6 +2642,21 @@ def run_job_inner(job: dict) -> None:
f"tier — the dev transformer needs more headroom than this Mac "
f"has. Bump to 64+ GB or render the longer clip in one shot."
)
+ # Y1.036 — Extend always loads the Q8 `transformer-dev.safetensors`
+ # weights (CFG-guided dev transformer, not the 8-step distilled).
+ # Pre-Y1.024 the Q4 dir incidentally shipped a copy of that file as
+ # download bloat, so Extend silently worked on Q4-only installs. The
+ # Y1.024 download filter pruned the dupe and exposed that Extend is
+ # structurally Q8-class. Gate it the same way Keyframe does and route
+ # the helper to the Q8 dir.
+ ext_missing = q8_missing_files()
+ if ext_missing:
+ raise RuntimeError(
+ f"Extend requires the full Q8 model at {Q8_LOCAL_PATH}. "
+ f"Missing {len(ext_missing)} file(s): {', '.join(ext_missing[:3])}"
+ f"{' …' if len(ext_missing) > 3 else ''}. "
+ f"Click \"Download Q8\" in Pinokio to install it (~37 GB)."
+ )
src = p["video_path"]
if not src or not Path(src).exists():
raise RuntimeError(f"source video for extend not found: {src}")
@@ -2667,6 +2695,10 @@ def run_job_inner(job: dict) -> None:
"action": "extend",
"id": job["id"],
"params": {
+ # Y1.036: explicit model_dir → Q8. Helper's get_pipe("extend")
+ # caches per-model-dir so this rebuilds the pipe only when
+ # the dir actually flips.
+ "model_dir": str(Q8_LOCAL_PATH),
"prompt": p["prompt"],
"negative_prompt": p.get("negative_prompt", ""),
"video_path": src,
@@ -2679,7 +2711,7 @@ def run_job_inner(job: dict) -> None:
},
}
push(f"Extend via helper: id={job['id']} src={Path(src).name} +{p['extend_frames']}f · "
- f"steps={steps} cfg={cfg_scale}")
+ f"steps={steps} cfg={cfg_scale} (Q8 dev transformer)")
result = HELPER.run(job_spec)
if "seed_used" in result:
push(f"seed used: {result['seed_used']}")
@@ -2690,7 +2722,11 @@ def run_job_inner(job: dict) -> None:
"params": {**p, "command": "extend"},
"started": job.get("started_at"),
"elapsed_sec": round(time.time() - job["started_ts"], 2) if job.get("started_ts") else None,
- "fps": FPS, "model": MODEL_ID, "queue_id": job["id"],
+ # Y1.036 — Extend always loads Q8 `transformer-dev` regardless of
+ # the panel's nominal MODEL_ID (which tracks the Q4 distilled
+ # used by T2V/I2V/Standard). Record the actual path so the
+ # /info modal and historical analyzers don't misreport.
+ "fps": FPS, "model": str(Q8_LOCAL_PATH), "queue_id": job["id"],
"helper_elapsed_sec": result.get("elapsed_sec"),
"output_codec": output_codec_settings(),
}
@@ -6878,17 +6914,21 @@ async function poll() {
if (document.getElementById('quality').value === 'high') setQuality('standard');
}
- // Keyframe (FFLF) requires Q8 — server enforces it (see run_job_inner). The
- // UI was previously silently downgrading the user to Standard when they
- // picked keyframe with Q8 missing, then the server would 500 on submit.
- // Disable Generate + show a clear reason while in that state.
+ // Keyframe (FFLF) and Extend both require Q8 — server enforces it (see
+ // run_job_inner). The UI was previously silently downgrading the user to
+ // Standard when they picked keyframe with Q8 missing, then the server
+ // would 500 on submit. Disable Generate + show a clear reason while in
+ // that state. Y1.036 added Extend to the same gate after the Y1.024
+ // download trim exposed that Extend is structurally Q8-class.
const genBtn = document.getElementById('genBtn');
- if (currentMode === 'keyframe' && !s.q8_available) {
+ const q8GatedMode = (currentMode === 'keyframe' || currentMode === 'extend');
+ if (q8GatedMode && !s.q8_available) {
genBtn.disabled = true;
+ const modeName = currentMode === 'keyframe' ? 'Keyframe (FFLF)' : 'Extend';
const left = (s.q8_missing || []).length;
genBtn.title = left > 0 && left < 6
- ? `Keyframe (FFLF) needs Q8 — ${left} file(s) still downloading.`
- : 'Keyframe (FFLF) needs the Q8 model. Click "Download Q8 (~37 GB)" in Pinokio.';
+ ? `${modeName} needs Q8 — ${left} file(s) still downloading.`
+ : `${modeName} needs the Q8 model. Click "Download Q8 (~37 GB)" in Pinokio.`;
genBtn.textContent = 'Generate · Q8 required';
} else if (genBtn.disabled && genBtn.textContent.startsWith('Generate · Q8')) {
// Restore — only do so if WE were the ones who disabled it, otherwise
@@ -7508,7 +7548,11 @@ function updateModelsCard(s) {
// Dismissible: a user who deliberately doesn't want Q8 (storage budget,
// they only do T2V Quick/Standard) can × this away and we'll respect it
// until either model state changes or they re-summon the modal.
+ // Y1.036 — Extend joins FFLF and High in needing Q8. The Extend pipeline
+ // loads `transformer-dev.safetensors` for CFG-guided denoise; Q4 doesn't
+ // ship it after the Y1.024 download trim, so surface the same CTA here.
const needsQ8 = (currentMode === 'keyframe')
+ || (currentMode === 'extend')
|| (document.getElementById('quality').value === 'high');
if (needsQ8 && !q8Ok && tier.allows_q8 !== false) {
if (dismissed) { card.style.display = 'none'; return; }
@@ -7516,6 +7560,7 @@ function updateModelsCard(s) {
card.classList.add('state-warn', 'dismissible');
icon.textContent = '⬇';
const reason = currentMode === 'keyframe' ? 'FFLF needs the Q8 model'
+ : currentMode === 'extend' ? 'Extend needs the Q8 model'
: 'High quality needs the Q8 model';
title.textContent = reason;
const missing = (s.q8_missing || []).length;
diff --git a/mlx_warm_helper.py b/mlx_warm_helper.py
index 01cbf21..877a682 100644
--- a/mlx_warm_helper.py
+++ b/mlx_warm_helper.py
@@ -317,13 +317,26 @@ def _attach_loras(pipe, loras: list[dict] | None) -> None:
pipe._pending_loras = pairs
-def get_pipe(kind: str, loras: list[dict] | None = None):
+_extend_model_dir: str | None = None
+
+
+def get_pipe(kind: str, loras: list[dict] | None = None,
+ model_dir: str | None = None):
"""kind in {'t2v','i2v','extend'}; loras is an optional list of
{path, strength} dicts. When the requested LoRA set differs from
the cached pipeline's, the pipeline is rebuilt — LoRA fusion is a
- one-shot weight transformation, not a runtime toggle."""
+ one-shot weight transformation, not a runtime toggle.
+
+ Y1.036 — `model_dir` overrides the helper-default LTX_MODEL env var on a
+ per-call basis. Used for Extend, which needs the Q8 `transformer-dev`
+ weights even on a Standard-tier render. Pre-Y1.024 the Q4 dir incidentally
+ carried a copy of `transformer-dev.safetensors` (download bloat) so Extend
+ silently loaded from there; the Y1.024 download filter pruned the dupe and
+ exposed that Extend is structurally Q8-class. Cached alongside the LoRA
+ fingerprint so a model_dir flip rebuilds the pipe."""
global _t2v_pipe, _i2v_pipe, _extend_pipe
global _t2v_lora_key, _i2v_lora_key, _extend_lora_key
+ global _extend_model_dir
from ltx_pipelines_mlx import TextToVideoPipeline, ImageToVideoPipeline, ExtendPipeline
fp = _lora_fingerprint(loras)
@@ -348,19 +361,24 @@ def get_pipe(kind: str, loras: list[dict] | None = None):
_i2v_lora_key = fp
return _i2v_pipe
if kind == "extend":
- if _extend_pipe is None or _extend_lora_key != fp:
- if _extend_pipe is not None and _extend_lora_key != fp:
+ ext_dir = model_dir or MODEL_ID
+ if (_extend_pipe is None
+ or _extend_lora_key != fp
+ or _extend_model_dir != ext_dir):
+ if _extend_pipe is not None:
+ why = "LoRA set changed" if _extend_lora_key != fp else "model_dir changed"
emit({"event": "log",
- "line": f"LoRA set changed; reloading Extend pipeline."})
+ "line": f"{why}; reloading Extend pipeline."})
_extend_pipe = None
emit({"event": "log",
- "line": "Loading Extend pipeline (heavier — uses dev transformer)..."})
+ "line": f"Loading Extend pipeline (heavier — uses dev transformer at {ext_dir})..."})
pipe = ExtendPipeline(
- model_dir=MODEL_ID, gemma_model_id=GEMMA_PATH, low_memory=LOW_MEMORY,
+ model_dir=ext_dir, gemma_model_id=GEMMA_PATH, low_memory=LOW_MEMORY,
)
_attach_loras(pipe, loras)
_extend_pipe = pipe
_extend_lora_key = fp
+ _extend_model_dir = ext_dir
return _extend_pipe
# t2v
if _t2v_pipe is None or _t2v_lora_key != fp:
@@ -1085,7 +1103,11 @@ for line in sys.__stdin__:
# Extend supports LoRAs via the same _pending_loras hook;
# the dev transformer picks them up at load time just like T2V/I2V.
loras = p.get("loras") or []
- pipe = get_pipe("extend", loras=loras)
+ # Y1.036 — Extend needs the Q8 `transformer-dev` weights. Panel
+ # passes the resolved Q8 path via params.model_dir; falling back
+ # to the helper's MODEL_ID is the legacy behavior.
+ ext_model_dir = p.get("model_dir")
+ pipe = get_pipe("extend", loras=loras, model_dir=ext_model_dir)
video_path = p["video_path"]
if not os.path.exists(video_path):
raise RuntimeError(f"source video not found: {video_path}")