RootAccess4Life/cond-dim-indextts-1.5

IndexTTS-1.5 — pinned mirror for cond-DIM

An unmodified, commit-pinned mirror of IndexTeam/IndexTTS-1.5, bundled with the matching index-tts source tree, so the cond-DIM paper's IndexTTS results stay reproducible.

⚠️ Licensing — read before use. Upstream IndexTeam/IndexTTS-1.5 is tagged apache-2.0 on the Hub, but the bundled index-tts-src/LICENSE is the bilibili Model Use License Agreement, which carries its own restrictions. These are not the same terms. Review both, and defer to upstream, before any use beyond research.

What this is

Upstream	IndexTeam/IndexTTS-1.5
Pinned commit	25851a6036dfd3095bb70fb3c8f49217104672c3
Modifications	None to the weights. Adds a vendored index-tts-src/ tree.
Purpose	Reproducibility anchor for the cond-DIM paper

This is the base checkpoint that cond-DIM edits — not an unlearned model.

Files

gpt.pth                     GPT backbone (1.17 GB)
bigvgan_generator.pth       BigVGAN vocoder generator (536 MB)
bigvgan_discriminator.pth   BigVGAN discriminator (1.65 GB, training only —
                            not needed for inference)
dvae.pth                    discrete VAE (243 MB)
bpe.model                   tokenizer
config.yaml                 model config
index-tts-src/              vendored upstream index-tts source tree

bigvgan_discriminator.pth is mirrored only for completeness with upstream; inference never loads it. Skip it to save ~1.65 GB:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="RootAccess4Life/cond-dim-indextts-1.5",
    local_dir="models/index_tts",
    ignore_patterns=["bigvgan_discriminator.pth", "index-tts-src/*"],
)

Then:

export CD_INDEXTTS_DIR=models/index_tts
python scripts/run_eval.py --config configs/indextts.yaml

IndexTTS needs its own isolated environment; the repo automates that with bash scripts/setup_indextts.sh.

Why IndexTTS-1.5 in this paper — and an honest caveat

IndexTTS-1.5 is architecturally the most distinct of the three backbones: autoregressive but perceiver-conditioned, with a BigVGAN vocoder. It is the strongest test of whether cond-DIM is genuinely host-independent rather than tuned to one conditioning design.

Erasure on IndexTTS is partial. BigVGAN carries its own speaker encoder, which is a second identity path that a conditioning-path edit does not touch. cond-DIM erases what flows through the conditioning path; residual identity can still leak through the vocoder. This is reported as a limitation in the paper rather than smoothed over.

Related

• 💻 Code: https://github.com/pujariaditya/cond-dim-tts-unlearning
• 🔇 Live demo: https://huggingface.co/spaces/RootAccess4Life/cond-dim-demo
• 🎧 Audio samples: https://huggingface.co/datasets/RootAccess4Life/cond-dim-samples
• 📦 XTTS-v2 mirror: https://huggingface.co/RootAccess4Life/cond-dim-xtts-v2

Citation

@misc{pujari_conddim,
  title  = {cond-DIM: One Method, Three Conditioning Paths for Host-Independent
            Speaker Unlearning in Autoregressive Zero-Shot TTS},
  author = {Pujari, Aditya and Rattani, Ajita},
  note   = {Preprint},
}

Please also cite IndexTTS upstream. All credit for these weights belongs to IndexTeam.