Lower GPU duration estimates and add 300s hard cap

Fix dynamic duration arg order mismatch + leftover torchaudio import

Standardise crossfade defaults to 2s/3dB across all three models

Dynamic GPU duration: reserve only what each run actually needs

Cleanup: remove duplication and dead code across app.py

Unify mono and stereo crossfade into a single _cf_join function

Refactor: consolidate segmentation and crossfade into shared helpers

Add crossfade Duration/Boost controls to MMAudio and HunyuanFoley tabs

Fix crossfade: use equal-power fade envelopes in all 3 models

Fix audio WAV header: ensure contiguous memory layout before torchaudio.save

HunyuanFoley: resample standalone audio output to 44100 Hz

Fix SyntaxError: move MMAUDIO_WINDOW constant above @spaces.GPU decorator

MMAudio: sliding-window segmentation for videos longer than 8 s; remove duration slider

HunyuanFoley: sliding-window segmentation for videos longer than 15 s

Fix CLAP from_pretrained daemonic process error in ZeroGPU worker

Pre-download CLAP model at startup to fix daemonic process error in ZeroGPU

Fix MMAudio inference_mode error (torch.no_grad); fix Hunyuan weights subdir path

.

Revert hunyuan checkpoint to .pth (truncated progress bar misled filename)

Add open_clip dep; remove tab descriptions; fix hunyuan .pt→.pth rename

Update UI descriptions: scene-type guidance instead of duration

Fix audiotools: install from git (PyPI version requires Python 3.11+)

Add torchdiffeq, audiotools, mmcv deps for MMAudio/Hunyuan/TARO

Add missing deps: colorlog (MMAudio) and loguru (HunyuanVideoFoley)

Fix mmaudio import: add MMAudio dir to sys.path in generate_mmaudio

Pin torchaudio==2.5.1+cu124 and add torch version diagnostic

Add torchaudio via PyTorch cu121 extra-index-url in requirements.txt

Fix torchaudio CUDA mismatch: install via PyTorch cu121 index in pre-install.sh

Fix torchaudio missing dep and TARO sys.path for bare imports

Upgrade to Gradio 5.25.0 to fix ZeroGPU localhost error

Fix: pin gradio==4.44.1 to fix ZeroGPU localhost ValueError

Fix launch() call for ZeroGPU — remove server_name override

Fix: call get_random_seed(), add server_name to launch() for ZeroGPU

Fix mmcv install: use mim via pre-install.sh instead of pip source build

Fix: move mmcv to requirements.txt, remove runtime install that caused infinite rebuild loop

Fix HunyuanFoley checkpoint path and add README metadata

Add HuggingFace Spaces metadata to README

Remove non-inference files from all three model folders

Add .gitignore and untrack MMAudio binary training examples

.

Restructure app.py: multi-model support (TARO, MMAudio, HunyuanFoley)

Show slot 0 by default; reveal extra slots on Generations slider drag

Rename 'Number of Samples' to 'Generations'

Multi-sample support, last-segment tail anchor fix, dynamic samples cap, gr.Blocks UI, duration=600

Remove share=True (redundant on HF Spaces)

Add crossfade duration/dB controls, inference caching, share=True

Support videos longer than 8.2s via overlapping inference + 2s +3dB crossfade stitching

Fix all indentation errors - consistent 4-space indentation throughout

Fix RuntimeError: pad onset_feats to truncate_onset=120 when video is shorter than 8.2s

Fix all indentation errors - consistent 4-space indentation throughout