idefinity/ming-3.0-tts-4step-distill-int4

ming-3.0-tts 4-step distilled DiT

A 4-step CFG-free flow DiT distilled from inclusionAI/Ming-3.0-tts-0.5B (teacher: 10-step, CFG=2).

Files

• distill/dit4.onnx — 4-step distilled DiT, fp32 ONNX.
• distill/dit4_int4.onnx — MatMulNBits 4-bit quantized DiT.
• distill/student_4step.pt — PyTorch student state dict.
• samples/teacher.wav, samples/student.wav, samples/student_int4.wav.
• metrics.json, diagram.png.

Metrics

Metric	fp32 student	int4 student
ASR CER	0.3976	0.4368
Speaker cosine	0.9340	0.9328
Log-mel L1	3.1799	2.9136
Duration ratio	0.996	0.996
DiT int4 probe cosine	—	0.992514

Training: 200 SFW Chinese utterances, 13,112 latent frames, 2,000 optimizer steps, batch 16, AdamW lr 1e-4, fp32.

Listening

• Teacher — 10-step CFG=2
• Student — 4-step CFG-free
• Student int4

Text: 雨落在窗台上，远处的灯一盏一盏亮起来。

Scope

This repository contains the distilled DiT and its int4 quantization. The LM, aggregator, encoder, and vocoder are not bundled here. The diagram compares 1/2/3/4-step students; this repository is the 4-step entry. SFW only. The base model is MIT licensed and its usage disclaimer applies.