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.

chore(panel): Train tab polish — rename, NEW badge, fix preset skew, drop dead v3 prose

mrbizarro4255f126e58b
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index 1d4accf..cedd88d 100644
--- a/mlx_ltx_panel.py
+++ b/mlx_ltx_panel.py
@@ -14836,7 +14836,7 @@ HTML = r"""<!doctype html>
<nav class="workflow-tabs" id="workflowTabs">
<button data-workflow="manual" class="active">Manual</button>
<button data-workflow="characters">Characters<span class="new-badge">NEW</span></button>
- <button data-workflow="train">Train</button>
+ <button data-workflow="train">Train Character<span class="new-badge">NEW</span></button>
</nav>
<!-- The original manual form is unchanged; it sits below in the DOM
and is shown/hidden by JS as the workflow tab toggles. -->
@@ -15582,14 +15582,13 @@ HTML = r"""<!doctype html>
</div>
<!-- Character LoRA strength. Default 0.8 — slightly under 1.0
- because Aria/Bizarro/Salo v2 LoRAs are mildly over-trained
- (5000 steps vs Lightricks-recommended 2000) and bake some
- training-time visual quirks (sparkles, occasional mesh).
- At 0.8 the identity still locks but the baked artifacts
- fade noticeably. The slider is exposed so power users can
- dial up to 1.0 for max identity or down to 0.5 to escape
- artifacts when prompts trigger them. v3 retrains with the
- corrected recipe should let us safely return to 1.0 default. -->
+ because the workshop/copper-pipe/gas-lamp prompt content
+ triggers mesh/sparkle artifacts in the BASE LTX-2.3 model
+ (confirmed via base-only-no-LoRA test 2026-05-17, see
+ lora-lab/STATE.md). At 0.8 the identity still locks and
+ the artifact-prone prompt regions render cleaner. Slider
+ exposed so users can dial up to 1.0 for max identity, or
+ down to 0.5 when a prompt triggers the base-model basin. -->
<div class="characters-strength-row">
<label for="charactersStrength" class="characters-strength-label">
Character LoRA strength
@@ -15830,7 +15829,7 @@ HTML = r"""<!doctype html>
<div class="pill-group cols-3" id="trainPresetGroup">
<button type="button" class="pill-btn active" data-train-preset="quick"><span>Quick <span class="rec-badge">Recommended</span></span><span class="sub" id="trainPresetQuickSub">~30 min · rank 8 · 512px</span></button>
<button type="button" class="pill-btn" data-train-preset="medium"><span>Medium</span><span class="sub" id="trainPresetMediumSub">~2 h · rank 16 · 576px</span></button>
- <button type="button" class="pill-btn" data-train-preset="high"><span>High</span><span class="sub" id="trainPresetHighSub">~4 h · rank 32 · 768px</span></button>
+ <button type="button" class="pill-btn" data-train-preset="high"><span>High</span><span class="sub" id="trainPresetHighSub">~2 h 50 min · rank 32 · 5000 steps · 512px</span></button>
</div>
<details class="train-advanced">
@@ -17313,8 +17312,8 @@ const TRAIN = {
label: 'Quick', subtitle: '~30 min · rank 8 · 512px' },
medium: { steps: 3000, rank: 16, resolution: 576, seconds_per_step: 2.2, ram_peak_gb: 18,
label: 'Medium', subtitle: '~2 h · rank 16 · 576px' },
- high: { steps: 5000, rank: 32, resolution: 768, seconds_per_step: 3.0, ram_peak_gb: 28,
- label: 'High', subtitle: '~4 h · rank 32 · 768px' },
+ high: { steps: 5000, rank: 32, resolution: 512, seconds_per_step: 2.0, ram_peak_gb: 28,
+ label: 'High', subtitle: '~2 h 50 min · rank 32 · 5000 steps · 512px' },
},
// Mirror of the server-side TRAIN_STYLE_PRESETS. Style table differs from
// character: quick uses rank 16 (not rank 8), and "high" adds steps not