1.0.8
SWivid/F5-TTS1.0.8Mar 24, 2025by SWivid
AI Summary
Addresses a critical training instability bug and improves Docker usage while adding options for model weight saving.
Key Highlights
- Fixed fatal bug in log_samples causing training instability
- Added option to save non-EMA model weights
- Improved Docker usage and documentation
New Features
- Training stability fix
- Non-EMA model saving option
- Docker improvements
Full Release Notes
**Full Changelog**: https://github.com/SWivid/F5-TTS/compare/1.0.3...1.0.8 - v1.0.8 fix a fatal bug with log_samples since 37eb3b50da8fdcb24f8b420b5f91246d85d4f5b5 that may be the root cause of training instability and unknown performance degradation - v1.0.7 add no_zero_init v1 variant path to SHARED.md; add option in finetune gradio to save non-ema model weight - v1.0.6 improves docker usage and along with other documentation - v1.0.5 update finetune_gradio.py for clearer guidance - v1.0.4 fix finetune_gradio.py vocab extend operation with checkpoint in .safetensors format