-
Mirror IndexTeam/IndexTTS-2 weights as safetensors
- gpt / s2mel / w2v_stats / feat1 / feat2 → .safetensors
- s2mel uses nested_v1 layout (keys flattened as net/<submodel>/<param>)
- feat1/feat2 use single-tensor layout (one key 'tensor')
- Other files (config.yaml, bpe.model, qwen0.6bemo4-merge/, glossary.yaml,
pinyin.vocab) copied verbatim from upstream.
- config.yaml weight references rewritten to .safetensors filenames.
Loaded by MAESTRO via the safetensors-aware torch.load shim in
backend/ai/models/index_tts2.py. -
initial commit
