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.
Python ai-filmmakingai-videoai-video-generatorapple-siliconhailuoimage-to-videolip-synclocal-ailora-trainingltx-2ltx-2-5ltx-videomacosminimax-h3mlxmusic-generationoffline-aipinokiotext-to-videovideo-generation
docs: refresh Extend Q8 release notes
mrbizarro3421a67911d9
diff --git a/README.md b/README.mdindex fa2cddc..76333ee 100644--- a/README.md+++ b/README.md@@ -147,7 +147,7 @@ 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 · ~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. |+| **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 and Extend. |The **Export** control can also run a lightweight ffmpeg Lanczos pass aftergeneration:@@ -404,17 +404,17 @@ ffmpeg -i in.mp4 \## Performance reference-Wall-clock times on an **M4 Mac Studio, 64 GB**:+Wall-clock times on an **M4-class Mac, 64 GB**:| Mode | Resolution | Frames | Steps | ~Time ||---|---|---|---|---|| T2V Quick | 640×480 | 121 (5s) | 8 | ~2 min |-| T2V Standard | 1280×704 | 121 (5s) | 8 | ~7 min |-| I2V Standard | 1280×704 | 121 (5s) | 8 | ~7 min |-| Extend (Q4 dev, cfg=1.0) | 768×416 | +6 latents (~2s) | 12 | ~12 min |-| Extend (Q4 dev, cfg=3.0 "Quality") | 1280×704 | +6 latents | 12 | ~30 min |+| T2V Balanced | 1024×576 → 1280×720 | 121 (5s) | 8 | ~5 min |+| T2V Standard | 1280×704 | 121 (5s) | 8 | ~8 min |+| I2V Standard | 1280×704 | 121 (5s) | 8 | ~9 min || High (Q8 two-stage) | 1280×704 | 121 | s1=15 + s2=3 | ~12 min |-| FFLF (Q8) | 768×416 | 121 | s1=15 + s2=3 | ~5 min |+| FFLF (Q8) | 768×416 | 121 | s1=15 + s2=3 | ~6 min |+| Extend (Q8 dev, cfg=1.0) | 768×416 | +3s | 12 | ~16 min |Experimental Speed smoke test on the same 64 GB machine, I2V Standard,768×512, 121 frames, 8 steps:diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.pyindex 79bf959..cfc2785 100644--- a/mlx_ltx_panel.py+++ b/mlx_ltx_panel.py@@ -1792,8 +1792,8 @@ def open_pinokio() -> None:# T2V/I2V Q8 HQ at 1280×704: ~50 GB peak, fits 64 GB cleanly# FFLF (Q8 dev) at 1280×704: > 64 GB peak, OOMs into swap# FFLF (Q8 dev) at 768×416: ~48 GB peak, fits 64 GB-# Extend (Q4 dev) at 1280×704: ~50 GB peak + 12 GB swap (240s/step)-# Extend (Q4 dev) at 768×416: ~46 GB peak, no swap (54s/step)+# Extend (Q8 dev) at 1280×704: > 64 GB peak, OOMs into swap+# Extend (Q8 dev) at 768×416: ~48 GB peak, fits 64 GB## `t2v_max_dim` / `i2v_max_dim` clamp T2V/I2V resolution. `keyframe_max_dim`# and `extend_max_dim` clamp those modes' resolutions specifically (since