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.

Y1.030 hide standalone Sharp upscaler

Salo0734017f2f86
diff --git a/README.md b/README.md
index a5d0385..81f3ab5 100644
--- a/README.md
+++ b/README.md
@@ -113,10 +113,10 @@ Exported/upscaled files use the same **Output format** setting as native
renders. Standard stays compact (`yuv420p` / CRF 18); Video production stays
lossless (`yuv444p` / CRF 0).
-The public export path uses ffmpeg Lanczos. Dev builds also expose the
-experimental LTX latent x2 upscaler for testing. Production can opt into it with
-`LTX_ENABLE_MODEL_UPSCALE=1`, but it is hidden by default because its
-doubled-latent peak can pressure 64 GB Macs.
+The public export path uses ffmpeg Lanczos. The experimental LTX latent x2
+upscaler can be forced with `LTX_ENABLE_MODEL_UPSCALE=1`, but it is hidden by
+default: the official LTX upscaler is meant to be followed by a second-stage
+denoise/refinement pass, and direct decode distorted faces in release tests.
Standard T2V/I2V also has an opt-in **Speed** control:
diff --git a/VERSION b/VERSION
index 9bcf9ec..82a27a9 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-Y1.028
+Y1.030
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index 834cec3..ba1d68c 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -119,11 +119,12 @@ def _detect_profile() -> str:
PROFILE = _detect_profile()
-# Sharp/LTX latent upscale is a dev-lab experiment. Production stays on the
-# safe ffmpeg Lanczos export path unless explicitly opted in; dev exposes it
-# by default so we can test the memory spike honestly before public release.
-_MODEL_UPSCALE_FLAG = _optional_bool_env("LTX_ENABLE_MODEL_UPSCALE")
-MODEL_UPSCALE_ENABLED = (PROFILE == "dev") if _MODEL_UPSCALE_FLAG is None else _MODEL_UPSCALE_FLAG
+# Sharp/LTX latent upscale is an unsafe lab experiment, not a release feature.
+# The official LTX pipelines use the latent upsampler only as the middle of a
+# two-stage flow: upscale latent -> denoise/refine at full resolution -> decode.
+# Running upsampler -> decode directly distorted faces and produced bad tail
+# frames in release tests, so keep it hidden unless explicitly opted in.
+MODEL_UPSCALE_ENABLED = _optional_bool_env("LTX_ENABLE_MODEL_UPSCALE") is True
# Default port: 8198 production, 8199 dev — so both panels can run side by
# side. LTX_PORT env var still overrides if the user wants something else.
DEFAULT_PORT = 8199 if PROFILE == "dev" else 8198
@@ -2529,7 +2530,7 @@ def make_job(form: dict[str, list[str]] | dict[str, str], *,
"model"
if MODEL_UPSCALE_ENABLED and f("upscale_method", "lanczos") == "model"
else "lanczos"
- ), # lanczos / model (model is dev/lab-only; see LTX_ENABLE_MODEL_UPSCALE)
+ ), # lanczos / model (model is lab-only; see LTX_ENABLE_MODEL_UPSCALE)
# LoRAs the user has enabled for this job. The UI submits a
# JSON-encoded array via the `loras` form field; we parse +
# validate here so the worker / helper layers receive a clean
@@ -2871,9 +2872,9 @@ def run_job_inner(job: dict) -> None:
"image": p["image"] if mode != "t2v" else None,
"loras": loras,
"accel": p.get("accel", "off"),
- # Public builds force this to Lanczos in make_job(); dev/lab
- # builds pass "model" so the helper can run the experimental
- # latent x2 upscaler.
+ # Normal builds force this to Lanczos in make_job(); only
+ # explicit lab opt-in passes "model" so the helper can run
+ # the experimental latent x2 upscaler.
"upscale": p.get("upscale", "off"),
"upscale_method": p.get("upscale_method", "lanczos"),
},
@@ -5756,10 +5757,11 @@ HTML = r"""<!doctype html>
= ffmpeg Lanczos resize, no detail recovery, near-instant.
Sharp = LTX 2.3 latent upscaler (~1 GB model), invents real
detail in latent space before VAE decode, ~30-60s slower.
- Sharp is enabled in dev/lab only after a release-test freeze
- showed the latent x2 peak can pressure 64 GB Macs too hard.
- Public builds use Fast by default unless opted in with
- LTX_ENABLE_MODEL_UPSCALE=1. -->
+ Sharp is hidden unless LTX_ENABLE_MODEL_UPSCALE=1 because
+ release tests showed the standalone latent x2 path can
+ distort faces and corrupt tail frames. The official LTX
+ contract is upsample + second-stage refinement, not direct
+ decode. -->
<div class="cz-control" id="upscaleMethodRow" style="display:none;">
<div class="cz-label">Method
<span class="cz-label-hint">how the upscale is done</span>