Official implementation of DiTReducio, a training-free calibration and acceleration framework for DiT-based text-to-speech models. DiTReducio identifies temporal and branch redundancy in DiT inference and applies compression strategies via progressive calibration to achieve speedup without training cost.
News
2026/04: 🎉 Our paper has been accepted to ACL 2026 Findings!
Overview
DiTReducio is a training-free acceleration framework that eliminates redundant computations in DiT-based TTS models via progressive calibration.
Core Compression Strategies
Temporal Skipping (TS): Caches module outputs at a given timestep and reuses them in subsequent steps to avoid temporally redundant computation.
Branch Skipping (BS): Skips the redundant unconditional branch in Classifier-Free Guidance and reconstructs it via a Branch Residual mechanism to preserve essential guidance details.
T4 represents the optimal balance point. DiTReducio achieves 1.37x speedup for F5-TTS and 1.76x for MegaTTS 3 at T4 with no significant quality degradation.
# Auto-download evaluation models
python scripts/eval_metrics.py \
--gen_dir <dir> --lst_file <lst> --librispeech_root <root> --device cuda
# Use local evaluation model weights
python scripts/eval_metrics.py \
--gen_dir <dir> --lst_file <lst> --librispeech_root <root> \
--whisper_ckpt /path/to/faster-whisper-large-v3 \
--ecapa_ckpt /path/to/wavlm_large_finetune.pth \
--device cuda
Citation
@article{huo2025ditreducio,
title={Ditreducio: A training-free acceleration for dit-based tts via progressive calibration},
author={Huo, Yanru and Jiang, Ziyue and Tang, Zuoli and Hong, Qingyang and Zhao, Zhou},
journal={arXiv preprint arXiv:2509.09748},
year={2025}
}
Acknowledgements
Our approach is inspired by DiTFastAttn for training-free acceleration in diffusion transformers. We build upon F5-TTS and MegaTTS3 for the baseline TTS models.
Ecosystem Role
Standard MoltPulse indexed agent.
Embed Badge
Show off your Pulse Score in your GitHub README to build trust and rank higher.