Load cached encoder before downloading remaining model

Stage persistent model components before local loading

Split persistent text encoder and runtime model caches

Persist Hugging Face model cache in mounted bucket

Process full uploaded video duration at 24 fps

Fix T4 startup CUDA memory exhaustion

Lazy-load gated LTX LoRAs with token diagnostics

Use mounted data bucket for render history

Add multi-effect LTX workflow

add aoti for speed up (#3)

Switch backend to native LTX-2 (ICLoraPipeline) (#2)

seed: write the used seed back into the Seed slider (drop separate 'Seed used' field)

point LoRA to official Lightricks LTX-2.3 Creative Lab repos

elaborate example prompts (more detail, README-aligned trigger/format preserved)

trim GPU duration estimate to ~2.3x measured runtime (quota efficiency)

raise default resolution/frames to training bucket (quality fix)

revert to plain track_tqdm progress (Z-Image pattern), drop generator/callback

randomize seed by default; move 'Seed used' into advanced settings

merge: keep AOTI speedup + native progress-bar generator

native progress bar via generator (forwards on ZeroGPU)

Use precompiled AOTI transformer blocks (ZeroGPU speedup) (#1)

concise description: base model + LoRA links, via diffusers

audio folded into single prompt; inpaint example None->placeholder mask

fixes: natural-speed/aspect, audio, progress, prompts, examples

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub