Model Updates 模型更新

This page tracks notable changes to the ONNX models and the inference pipeline shipped with SXSEditor, grouped by model version. For application-level changes, see Application Updates.

本页按模型版本追踪 SXSEditor 内置 ONNX 模型与推理管线的重大变更。应用层变更请见应用更新。

ℹ️

Model files are not stored in the Git repository (they are git-ignored). They are downloaded automatically on first launch, or placed manually in onnx_models/. See FAQ for download troubleshooting.

模型文件不包含在 Git 仓库中（已被 gitignore）。首次启动时会自动下载，也可手动放入 onnx_models/。下载问题请见 FAQ。

v1 (Current) v1（当前版本）

Re-optimized models for GPU — the ONNX models have been re-optimized for the DirectML GPU backend. Vocoder inference speed is now up to 6× faster.为 GPU 重新优化模型——ONNX 模型已针对 DirectML GPU 后端重新优化。声码器推理速度现在最高提升 6 倍。
ONNX opset upgraded to 20 — all models are now exported at opset 20 (the maximum allowed for DirectML compatibility), improving GPU kernel selection.ONNX opset 升级到 20——所有模型现在以 opset 20 导出（DirectML 兼容的最大值），改善 GPU 内核选择。
W16A32 quantization — added a W16A32 precision (FP16 weights + FP32 activations) that delivers near-FP32 quality with lower memory usage, resolving FP16 underflow issues in low-energy consonants.W16A32 量化——新增 W16A32 精度（FP16 权重 + FP32 激活），在降低显存占用的同时接近 FP32 质量，解决低能量辅音的 FP16 下溢问题。
SiFiGAN vocoder option — a second vocoder (SiFiGAN) is now available alongside the default Vocos-style vocoder, with automatic mel frame-rate adaptation.SiFiGAN 声码器选项——在默认 Vocos 风格声码器之外新增 SiFiGAN 声码器，自动适配 mel 帧率。
mel_transform precision boost — the mel spectrogram frontend now uses a Cooley-Tukey FFT implementation, improving signal-to-noise ratio from 49.86 dB to 83.54 dB.mel_transform 精度提升——mel 频谱前端改用 Cooley-Tukey FFT 实现，信噪比从 49.86 dB 提升到 83.54 dB。
Diffusion model GPU stability — fixed NaN output from the diffusion model on the DirectML backend by precomputing position-encoding tables, enabling stable GPU inference.扩散模型 GPU 稳定性——通过预算位置编码表修复了扩散模型在 DirectML 后端的 NaN 输出，实现稳定的 GPU 推理。
Fixed Vocos vocoder output explosion — resolved a reshaping bug in the vocoder's overlap-add stage that caused output explosion (electrical noise, wavRMS=167). PyTorch-vs-ONNX SNR improved from −75.35 dB to 84.24 dB.修复 Vocos 声码器输出爆炸——修复了声码器 overlap-add 阶段的 reshape 错误，该错误导致输出爆炸（电流声，wavRMS=167）。PyTorch 与 ONNX 的 SNR 从 −75.35 dB 提升到 84.24 dB。
FCPE pitch detector — a new ONNX-based F0 extraction model (Fast Context-Free Pitch Estimator) has been added. FCPE is now the default MIDI extraction tool, replacing Basic Pitch as the recommended option. It offers faster and more accurate F0-to-note conversion. Auto-downloaded from ModelScope. Registered as a new model group in modelRegistry.js.FCPE 音高检测器——新增基于 ONNX 的 F0 提取模型（Fast Context-Free Pitch Estimator）。FCPE 现在是默认 MIDI 提取工具，取代 Basic Pitch 成为推荐选项。它提供更快、更准确的 F0 转音符转换。从 ModelScope 自动下载。在 modelRegistry.js 中注册为新的模型组。
ROSVOT voice onset detection — a new ONNX-based voice onset detection model has been added. ROSVOT detects the precise onset timing of vocal segments, improving note boundary alignment during MIDI extraction from audio. Auto-downloaded from ModelScope and registered as a new model group.ROSVOT 发声起始检测——新增基于 ONNX 的发声起始检测模型。ROSVOT 检测人声片段的精确起始时间，改善从音频提取 MIDI 时的音符边界对齐。从 ModelScope 自动下载并注册为新的模型组。
QDIT int8 diff_step — QDIT-quantized int8 diff_step models are now supported, using a new ONNX signature with improved quantization-aware inference for better quality at int8 precision. Legacy int8 diff_step models are automatically detected on load — if a legacy model is found, a prompt suggests switching to the QDIT version for better synthesis quality.QDIT int8 diff_step——现在支持 QDIT 量化的 int8 diff_step 模型，使用新的 ONNX 签名和改进的量化感知推理，在 int8 精度下提供更好音质。加载时自动检测旧版 int8 diff_step 模型——如果检测到旧版模型，会提示切换到 QDIT 版本以获得更好的合成质量。

v0 (Deprecated) v0（已弃用）

The original model release. It is no longer provided and has been fully superseded by v1. Known limitations that led to its deprecation:

最初的模型版本。不再提供，已被 v1 完全取代。导致其弃用的已知限制：

Exported at ONNX opset 17/18, without GPU-specific optimization; vocoder inference was slow.以 ONNX opset 17/18 导出，未针对 GPU 优化；声码器推理速度慢。
Static input shapes limited reference-audio length; longer references fell back to a JS implementation that produced NaN ptMelData.静态输入形状限制了参考音频长度；更长参考会回退到产生 NaN ptMelData 的 JS 实现。
The diffusion model produced NaN output on the DirectML backend for large dynamic shapes.扩散模型在大动态形状下会在 DirectML 后端产生 NaN 输出。
The Vocos vocoder contained an overlap-add reshaping bug causing output explosion (electrical noise).Vocos 声码器存在 overlap-add reshape 错误，导致输出爆炸（电流声）。
FP8 precision was incompatible with the DirectML execution provider.FP8 精度与 DirectML 执行提供者不兼容。