Auto-detect available ffmpeg codec (libx264/libopenh264/mpeg4 fallback)

Remove -crf and complex vf filter - use simple -pix_fmt yuv420p

Fix ffmpeg: use shell=True, reorder -crf option

Fix ffmpeg: use subprocess.run with error checking, absolute temp path

Add debug listing to download endpoint

Add custom /api/download endpoint to bypass Gradio file serving 403

Add debug logging for file access

Debug: allowed_paths=['/'] + logging

Add allowed_paths /tmp for file serving

Set GRADIO_TEMP_DIR=/tmp for file serving

Save results to Gradio cache dir for proper file serving

Save results to temp dir for Gradio 5 file serving

Fix allowed_paths with absolute path for results directory

Add allowed_paths for results directory

fix: use direct paste-back instead of get_image (requires face parsing model we don't init)

fix: use proper tensor dtype conversion in inference loop (FloatTensor fails on CUDA half tensors)

Fix gradio_client schema bug + update app

fix: correct get_whisper_chunk parameter order (device was getting librosa_length)

fix: update model loading API

fix: revert to eager model loading — L4 GPU loads models fast, deferred loading caused inference errors

fix: defer model loading to first inference — unblocks Gradio startup

fix: remove debug=True from launch

fix: resolve NameError, align model paths to models/, remove @spaces.GPU

Update app.py

Update app.py

Update app.py

Upload 68 files