fix: remove enable_model_cpu_offload, use pipe.to(cuda) on xlarge for 22B model

Upload app.py with huggingface_hub

use enable_model_cpu_offload instead of pipe.to(cuda) for 48GB VRAM

load model at module level so download happens during APP_STARTING

add peft for LoRA loading

fix: use /tmp for HF cache (no bucket mounted)

fix: duration estimator kwargs handling

refactor: remove fal.ai, run LTX-2.3 3DREAL locally on ZeroGPU with diffusers

Make layout fully responsive — stack columns on mobile, fix horizontal overflow

Make layout fully responsive — stack columns on mobile, remove min_width overflow

Link LTX 3DREAL model in Space metadata

Use Gradio examples only for cached demos

Remove redundant examples loader and video borders

Fix Gradio media routes

Fix Gradio 6 launch styling

Initial LTX 3DREAL fal demo

initial commit