Settings 设置
Open Settings from the menu bar: SXSEditor > Settings. The Settings window uses a sidebar layout with categories: General, Inference, Audio, and Model. Most settings apply immediately on change (auto-saved with a 300ms debounce); language changes require an app restart.
从菜单栏打开设置：SXSEditor > Settings。设置窗口采用侧边栏布局，分类包括：通用、推理、音频和模型。大多数设置在更改后立即生效（300ms 防抖自动保存）；语言更改需要重启应用。
General 通用
Language 语言
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Language界面语言 | zh-CN (中文简体) or en (English). Requires app restart to take effect.zh-CN（中文简体）或 en（English）。需要重启应用生效。 |
System locale系统区域设置 |
Theme 主题
Select from built-in and user themes. Changes apply immediately to all windows (hot-swap, no restart needed). Built-in themes:
从内置和用户主题中选择。更改立即应用到所有窗口（热切换，无需重启）。内置主题：
| IDID | Name名称 | Description说明 |
|---|---|---|
dark-aurora |
Aurora DarkAurora Dark | Default dark theme with blue-purple accents默认深色主题，蓝紫色调 |
light-paper |
Paper LightPaper Light | Light theme with white background浅色主题，白色背景 |
midnight-amber |
Midnight AmberMidnight Amber | Dark theme with warm amber accents深色主题，暖琥珀色调 |
acg |
ACGACG | Anime/game-inspired color scheme动漫/游戏风格配色 |
Theme actions: Edit Current Theme (visual editor with 20-step undo/redo), Import Theme (.theme.json), Export Theme, Delete User Theme, Reset to Default. User themes are stored at %APPDATA%\sxseditor\themes\<theme-id>.theme.json.
主题操作：Edit Current Theme（可视化编辑器，支持 20 步撤销/重做）、Import Theme（.theme.json）、Export Theme、Delete User Theme、Reset to Default。用户主题存储于 %APPDATA%\sxseditor\themes\<theme-id>.theme.json。
Inference 推理
Inference Hardware 推理硬件
Configure which hardware devices are used for neural network inference.
配置用于神经网络推理的硬件设备。
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Inference Provider推理提供者 | ORTNODE (DirectML/CPU, main process) or ORTWEB (WebNN NPU/GPU, renderer). ORTNODE shows local GPU/CPU; ORTWEB shows WebNN NPU/GPU.ORTNODE（DirectML/CPU，主进程）或 ORTWEB（WebNN NPU/GPU，渲染进程）。ORTNODE 显示本地 GPU/CPU；ORTWEB 显示 WebNN NPU/GPU。 |
ORTNODE |
| Device Mode设备模式 | Smart (auto-select best device, prefer discrete GPU), Manual (specify single device for all models), Advanced (assign different devices per model group: SVS Diffusion, SVS Encoder, SVS Auxiliary, RMVPE, RosVot)Smart（自动选择最佳设备，优先独显）、Manual（为所有模型指定单一设备）、Advanced（为每个模型组单独指定设备：SVS Diffusion、SVS Encoder、SVS Auxiliary、RMVPE、RosVot） | SmartSmart |
| Inference Hardware推理硬件 | Auto-select (prefer discrete GPU) or a specific device. Available devices: Discrete GPU, Integrated GPU (via DirectML), NPU (via WebNN), CPU (fallback).自动选择（优先独显）或指定设备。可用设备：独立显卡、集成显卡（通过 DirectML）、NPU（通过 WebNN）、CPU（回退）。 | AutoAuto |
The settings page shows real-time status of WebNN, NPU, and GPU availability. Hardware detection runs once after app startup and caches results (no re-probing during synthesis).
设置页会实时显示 WebNN、NPU 和 GPU 的可用状态。硬件检测在应用启动后仅运行一次并缓存结果（合成时不再重复探测）。
Preview Inference Parameters 预览推理参数
Used when playing back in the editor (fast preview).
在编辑器中播放预览时使用（快速预览）。
| Parameter参数 | Default默认 | Range / Step范围 / 步进 | Description说明 |
|---|---|---|---|
| Diffusion Steps扩散步数 | 1616 | 4–64 / step 44–64 / 步进 4 | Fewer steps = faster, lower quality. Recommended preview: 8–16.步数越少越快但音质下降。预览推荐：8–16。 |
| CFG StrengthCFG 引导强度 | 3.03.0 | 0–10 / step 0.50–10 / 步进 0.5 | Higher = more aligned with conditions. 0 = skip unconditional prediction (2x speed). Recommended 1.0–5.0.值越大越贴合条件。0 = 跳过无条件预测（速度翻倍）。推荐 1.0–5.0。 |
| CFG RescaleCFG Rescale 系数 | 0.750.75 | 0–1 / step 0.050–1 / 步进 0.05 | Mitigates over-guidance artifacts. Recommended 0.5–0.9.缓解过度引导伪影。推荐 0.5–0.9。 |
| Sampler求解器 | EulerEuler | Euler / Heun / Extrapolated Euler / STORK-2Euler / Heun / Extrapolated Euler / STORK-2 | Diffusion sampling solver. See Diffusion Sampler below.扩散采样求解器。见下方 扩散求解器。 |
Export Inference Parameters 导出推理参数
Used when exporting WAV files (high quality).
导出 WAV 文件时使用（高质量）。
| Parameter参数 | Default默认 | Range / Step范围 / 步进 | Description说明 |
|---|---|---|---|
| Diffusion Steps扩散步数 | 3232 | 4–64 / step 44–64 / 步进 4 | More steps = higher quality. Recommended export: 32.步数越多音质越高。导出推荐：32。 |
| CFG StrengthCFG 引导强度 | 3.03.0 | 0–10 / step 0.50–10 / 步进 0.5 | Same as preview. Keep higher for export quality.同预览。导出时建议保持较高值。 |
| CFG RescaleCFG Rescale 系数 | 0.750.75 | 0–1 / step 0.050–1 / 步进 0.05 | Same as preview.同预览。 |
| Sampler求解器 | EulerEuler | Euler / Heun / Extrapolated Euler / STORK-2Euler / Heun / Extrapolated Euler / STORK-2 | Diffusion sampling solver. See Diffusion Sampler below. The export dialog also lets you override this per export.扩散采样求解器。见下方 扩散求解器。导出对话框也支持每次导出单独覆盖。 |
Diffusion Sampler 扩散求解器
The diffusion model iteratively denoises a mel spectrogram by integrating the flow-matching velocity field ODE. The sampler decides how each denoising step combines model evaluations into the state update. SXSEditor exposes four samplers, selectable independently for preview and export:
扩散模型通过积分 flow-matching 速度场 ODE 迭代去噪 mel 频谱。求解器决定每个去噪步如何将模型推理组合为状态更新。SXSEditor 提供四种求解器，预览与导出可独立选择：
| Sampler求解器 | NFE / step每步推理次数 | Description说明 |
|---|---|---|
| Euler (default)Euler（默认） | 11 | First-order explicit Euler, midpoint time evaluation. The original baseline implementation; fastest and most predictable.一阶显式欧拉法，中点时间评估。原始基线实现，最快且最稳定。 |
| HeunHeun | 22 | Second-order improved Euler (trapezoidal rule). Higher accuracy at the cost of roughly 2× inference time. Falls back to Euler on the final step to avoid evaluating t > 1.二阶改进欧拉（梯形公式）。精度更高，代价是推理时间约 2 倍。末步退化为 Euler 以避免 t > 1 越界。 |
| Extrapolated EulerExtrapolated Euler | 11 | Velocity-extrapolation heuristic inspired by STORK (ICLR 2026). Reuses the previous step's velocity to linearly extrapolate the next velocity — 1 inference/step. Not the full stabilized Runge-Kutta formulation; benefit is heuristic and strongest when the velocity field changes smoothly. Includes numerical-stability guards (amplitude / sign-flip / NaN) that fall back to Euler when extrapolation is unsafe.受 STORK（ICLR 2026）启发的速度外推启发式。复用上一步速度线性外推下一步速度——每步 1 次推理。并非论文完整的 stabilized Runge-Kutta 格式；收益为启发式，速度场缓变时效果最佳。含数值稳定性保护（幅度 / 符号翻转 / NaN），外推不安全时退化为 Euler。 |
| STORK-2STORK-2 | 11 | Paper-faithful implementation of STORK-2 (Tan et al., ICLR 2026, arXiv:2505.24210). Runge-Kutta-Gegenbauer 2nd-order recurrence with 8 sub-stages and Taylor-expansion virtual NFE — only 1 model inference per step despite the 8 sub-stages. Designed for stiff ODEs with an extended stability region (~2s² = 128×). Tradeoff: higher per-step algebraic cost (8 full-array passes) than Euler.STORK-2 论文原版实现（Tan et al., ICLR 2026, arXiv:2505.24210）。Runge-Kutta-Gegenbauer 二阶递推 + 8 个 sub-stage + Taylor 展开 virtual NFE——尽管 8 个 sub-stage，每步仅 1 次模型推理。专为刚性 ODE 设计，稳定性域扩展约 2s²=128 倍。代价：每步代数运算开销高于 Euler（8 趟全数组遍历）。 |
Both Extrapolated Euler and STORK-2 keep cross-step velocity history. In chunked preview inference (long fragments split into chunks), this history resets at every chunk boundary, which reduces their advantage — for chunked previews, Euler or Heun is usually the safer choice. The sampler setting is stored in previewSampler / exportSampler; the legacy value stork is silently mapped to extrap for backward compatibility.
Extrapolated Euler 与 STORK-2 都需要跨步速度历史。在分块预览推理（长片段被切成多块）时，每块边界都会重置历史，使其优势减弱——分块预览通常 Euler 或 Heun 更稳妥。求解器设置存储于 previewSampler / exportSampler；旧值 stork 会被静默映射为 extrap 以保持向后兼容。
Vocoder Chunk Settings Vocoder 分片设置
For long audio synthesis, the vocoder processes audio in chunks then crossfade-stitches them. Larger chunks use more VRAM but fewer stitch operations; smaller chunks save VRAM but increase stitch overhead.
长音频合成时，vocoder 按分片逐段推理再交叉淡入淡出拼接。分片越大显存占用越高但拼接次数越少；分片越小越省显存但拼接开销增加。
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Chunk Mode分片模式 | Smart (auto-allocate from VRAM budget) or Manual (specify frames directly)Smart（按显存预算自动分配）或 Manual（直接指定帧数） | SmartSmart |
| Chunk Frames (manual)分片帧数（手动） | Must be a multiple of 8. Range 256–2048. ~20 seconds of audio at 1008 frames.必须是 8 的倍数。范围 256–2048。1008 帧约 20 秒音频。 | 10081008 |
Smart mode calculates VRAM budget = (VRAM − resident weights − diff_step activations ~2GB − OS reserve ~1GB) × 0.7 safety factor. The settings page shows a VRAM reference table (2/3/4/6/8/10/12/16/20/24GB) computed with the current precision and vocoder type, with the current GPU's tier highlighted.
智能模式计算显存预算 =（VRAM − 常驻权重 − diff_step 激活 ~2GB − OS 占用 ~1GB）× 0.7 安全系数。设置页提供显存对照表（2/3/4/6/8/10/12/16/20/24GB），按当前精度与 vocoder 类型实时计算，当前显卡对应行高亮。
Dynamic Thresholding 动态阈值
Dynamic thresholding clips extreme values in the diffusion latent space at a specified percentile, reducing artifacts from over-sharp predictions. Available for both preview and export.
动态阈值按指定百分位裁剪扩散潜空间中的极端值，减少过度尖锐预测导致的伪影。预览和导出均可用。
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Enable Dynamic Thresholding启用动态阈值 | Toggle dynamic thresholding on/off for the diffusion pass.为扩散过程开启/关闭动态阈值。 | Off关闭 |
| Percentile百分位 | Slider 0.9–0.999. Higher values clip more aggressively.滑块 0.9–0.999。值越高裁剪越激进。 | 0.9990.999 |
CFG Schedule CFG 调度
The CFG schedule controls how CFG strength varies across diffusion steps. Available for both preview and export.
CFG 调度控制 CFG 强度在扩散步间的变化方式。预览和导出均可用。
| Mode模式 | Description说明 |
|---|---|
| Constant (default)恒定（默认） | Same CFG strength throughout all diffusion steps.所有扩散步使用相同的 CFG 强度。 |
| Linear线性 | CFG strength ramps from high to low across steps.CFG 强度从高到低线性渐变。 |
| Cosine余弦 | CFG strength follows a smooth cosine curve across steps.CFG 强度按平滑余弦曲线变化。 |
| Custom自定义 | User-defined per-step CFG values.用户逐步指定 CFG 值。 |
Vocoder Overlap Frames Vocoder 重叠帧数
Controls the number of overlap frames used when stitching vocoder chunks. More overlap = smoother transitions but slightly longer processing; less overlap = faster but may introduce stitch artifacts.
控制拼接 vocoder 分片时使用的重叠帧数。重叠越多 = 过渡越平滑但处理时间略增；重叠越少 = 更快但可能引入拼接伪影。
| Option选项 | Range范围 | Default默认 |
|---|---|---|
| Vocoder Overlap FramesVocoder 重叠帧数 | 8–96 (must be a multiple of 8)8–96（必须是 8 的倍数） | 3232 |
Japanese Vocalization 日语发声方式
Japanese lyrics can be synthesized on the base multilingual model (no extra download) or via a dedicated JP LoRA model (in development). Three modes are available:
日语歌词可在基础多语言模型上合成（无需额外下载），或通过专用日语 LoRA 模型合成（开发中）。提供三种模式：
| Mode模式 | Description说明 | Download下载 |
|---|---|---|
| English Phoneme Migration英语音素迁移 | Maps Japanese kana to the nearest English ARPAbet phonemes. Uses the base multilingual model. ら-row maps to R, お-column maps to OW (diphthong).将日文假名映射为最接近的英语 ARPAbet 音素，在基础多语言模型上合成。ら行映射为 R，お段映射为 OW（双元音）。 | No否 |
| Hybrid Phonemes (default)混合音素（默认） | Builds on English Phoneme Migration by picking, for each kana, whichever is phonetically closer between English ARPAbet (phoneme-level) and Cantonese yue_ (syllable-level). Consonants prefer Cantonese (e.g. か-row uses Cantonese g = unaspirated /k/, closer to Japanese k than English aspirated K /kʰ/); pure vowels prefer Cantonese (e.g. あ→yue_aa1, お→yue_o1, purer than English AA/AO); ら-row uses Cantonese l /l/ to approximate the Japanese tap /ɾ/. Yōon, voiced rows, and kanji fall back to ARPAbet. No extra download.在英语音素迁移的基础上，对每个假名在英语 ARPAbet（音素级）与粤语 yue_（音节级）之间挑选发音最接近的一项。辅音优先粤语（如 か 行用粤 g 不送气 /k/，比英语 K 送气 /kʰ/ 更接近日语）；纯元音优先粤语（如 あ→yue_aa1、お→yue_o1，比英语 AA/AO 更纯净）；ら行用粤 l /l/ 接近日语弹音 /ɾ/。拗音、浊行、汉字回退 ARPAbet。无需额外下载。 | No否 |
| Japanese LoRA Model日语 LoRA 模型 | Dedicated Japanese fine-tuned model with native jp_ phonemes for best quality.专用日语微调模型，使用原生 jp_ 音素，音质最佳。 | Required (in development)需下载（开发中） |
Both English Phoneme Migration and Hybrid Phonemes use the same base multilingual model — the only difference is how each kana is mapped (English-only ARPAbet vs. ARPAbet/Cantonese per-kana selection). Hybrid mode is recommended for more natural-sounding Japanese. The phoneme boundary editor (per-note timing/volume) works the same in all modes.
英语音素迁移与混合音素均使用同一基础多语言模型——唯一区别在于每个假名的映射方式（纯 ARPAbet vs. ARPAbet/粤语按假名逐个挑选）。推荐使用混合音素以获得更自然的日语发音。音素边界编辑器（逐音符时长/音量调整）在所有模式下工作方式相同。
Audio 音频
Output Settings 输出设置
| Option选项 | Choices可选项 | Default默认 |
|---|---|---|
| Output Mode输出模式 | Shared (WASAPI Shared, other apps can play audio) / Exclusive (WASAPI Exclusive, lower latency 1–3ms, bit-perfect, blocks other apps)Shared（WASAPI 共享，其他应用可同时播放）/ Exclusive（WASAPI 独占，延迟更低 1–3ms，位完美，会阻止其他应用） | SharedShared |
| Output Device输出设备 | System Default or a specific audio device. Exclusive mode only supports WASAPI devices.系统默认或指定音频设备。独占模式仅支持 WASAPI 设备。 | System Default系统默认 |
| Sample Rate采样率 | 22050, 24000 (native), 44100 (CD), 48000 (standard), 96000 (high), 192000 (studio) Hz. The default is now 48000 Hz (was 24000 Hz); the model still operates internally at 24 kHz and the output is linearly upsampled.22050、24000（原生）、44100（CD）、48000（标准）、96000（高）、192000（录音棚）Hz。默认现在为 48000 Hz（原为 24000 Hz）；模型仍以 24 kHz 内部运行，输出经线性上采样。 | 48000 Hz48000 Hz |
| Export Sample Rate导出采样率 | 24000, 44100, 48000, 96000 Hz. Selectable per export. Linear upsampling from the 24 kHz model rate.24000、44100、48000、96000 Hz。每次导出可选。从 24 kHz 模型率线性上采样。 | 48000 Hz48000 Hz |
| Bit Depth位深度 | 32-bit Float (recommended), 32-bit Integer, 24-bit Integer, 16-bit Integer. Only applies in exclusive mode.32-bit Float（推荐）、32-bit Integer、24-bit Integer、16-bit Integer。仅独占模式生效。 | 32-bit Float32-bit Float |
| Buffer Size缓冲区大小 | 64 (~1.3ms), 128 (~2.7ms), 256 (~5.3ms), 512 (~10.7ms), 1024 (~21.3ms), 2048 (~42.7ms), 4096 (~85.3ms) samples. Smaller = lower latency but higher CPU/glitch risk.64（~1.3ms）、128（~2.7ms）、256（~5.3ms）、512（~10.7ms）、1024（~21.3ms）、2048（~42.7ms）、4096（~85.3ms）采样。越小延迟越低但 CPU 负载和爆音风险越高。 | 10241024 |
| Master Volume主音量 | Slider 0%–100%.滑块 0%–100%。 | 100%100% |
In exclusive mode, if the device doesn't support the selected sample rate/bit depth, SXSEditor automatically falls back to shared mode. Recommended buffer size in exclusive mode: 256 or below.
独占模式下，如果设备不支持所选采样率/位深度，SXSEditor 会自动回退到共享模式。独占模式推荐缓冲区大小：256 或更低。
Accompaniment tracks — per-track volume control is available for accompaniment tracks in the track manager. Accompaniment audio (WAV/MP3/FLAC/OGG/M4A/AAC) can be imported alongside vocal fragments and is mixed into the final export.
伴奏音轨——轨道管理器中支持伴奏音轨的逐轨音量控制。伴奏音频（WAV/MP3/FLAC/OGG/M4A/AAC）可与人声片段一起导入，并混入最终导出。
MIDI Extraction Tool MIDI 提取工具
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| MIDI Extract ToolMIDI 提取工具 | FCPE (default, recommended, ONNX-based F0 extraction) / Basic Pitch (neural network-based, stable) / RMVPE (F0-to-notes conversion, experimental, may produce suboptimal results)FCPE（默认，推荐，基于 ONNX 的 F0 提取）/ Basic Pitch（神经网络，稳定）/ RMVPE（F0 转音符，实验性，结果可能不理想） | FCPEFCPE |
FCPE Pitch Detector Settings FCPE 音高检测器设置
FCPE (Fast Context-Free Pitch Estimator) is now the default MIDI extraction tool. When FCPE is selected as the MIDI extraction tool, a dedicated settings group is shown (auto-hidden for other tools). The following global parameters control FCPE behavior:
FCPE（Fast Context-Free Pitch Estimator）现在是默认的 MIDI 提取工具。当选择 FCPE 作为 MIDI 提取工具时，会显示专用设置组（选择其他工具时自动隐藏）。以下全局参数控制 FCPE 行为：
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Threshold阈值 | Confidence threshold for pitch detection. Higher values reject uncertain F0 estimates.音高检测的置信度阈值。值越高，不确定的 F0 估计会被拒绝。 | 0.030.03 |
| Smoothing平滑 | Applies a smoothing filter to the extracted F0 curve before note conversion.在转换为音符前对提取的 F0 曲线施加平滑滤波。 | Off关闭 |
| Quantization Mode量化模式 | Strict (snap each note to the nearest semitone) or Pitch-Bend Preserving (keep F0 bend data as pitch curve instead of quantizing).严格（将每个音符吸附到最近的半音）或 保留弯音（保留 F0 弯音数据作为音高曲线而不量化）。 | Strict严格 |
| F0 RangeF0 范围 | Minimum and maximum F0 frequency for detection (Hz).检测的最小和最大 F0 频率（Hz）。 | 50–1100 Hz50–1100 Hz |
| Minimum Note Duration最短音符时长 | Shortest note length to emit (ms). Shorter F0 segments are merged into neighbors.输出的最短音符长度（ms）。更短的 F0 片段会被合并到相邻音符。 | 100 ms100 ms |
| Normalization归一化 | Enable/disable F0 normalization before note conversion.在转换为音符前启用/禁用 F0 归一化。 | Off关闭 |
Model 模型
Model Precision 模型精度
| Precision精度 | Description说明 | Disk Size磁盘占用 |
|---|---|---|
| FP16 (default)FP16（默认） | Half precision. Best balance of quality and speed for discrete GPUs (4GB+ VRAM).半精度。独显（4GB+ 显存）质量与速度的最佳平衡。 | ~1.5 GB约 1.5 GB |
| FP32FP32 | Full precision. Highest quality, very high VRAM usage. Requires 8GB+ VRAM.全精度。最高音质，显存占用很高。需要 8GB+ 显存。 | ~3 GB约 3 GB |
| INT8INT8 | Quantized. Slight quality loss, low VRAM. For integrated GPU or <2GB VRAM.量化。轻微音质损失，低显存。适用于集成显卡或 <2GB 显存。 | ~1 GB约 1 GB |
| INT8-NPUINT8-NPU | NPU-optimized fixed-dimension quantized. For NPU hardware.NPU 优化固定维度量化。适用于 NPU 硬件。 | ~1 GB约 1 GB |
Different precisions are stored independently and coexist — each has its own subdirectory. Switching does not require re-downloading. The model status list shows ✅ Ready (all files present) or ❌ N files missing for each precision.
不同精度独立存储且可共存——每种精度有独立子目录。切换无需重新下载。模型状态列表为每种精度显示 ✅ Ready（所有文件就绪）或 ❌ N files missing（缺失 N 个文件）。
QDIT int8 detection — when using int8 precision, the app automatically detects whether the diff_step model is a legacy int8 model or a QDIT-quantized model. If a legacy int8 model is detected, a prompt suggests switching to QDIT for better quality. QDIT models use a new ONNX signature with improved quantization-aware inference.
QDIT int8 检测——使用 int8 精度时，应用会自动检测 diff_step 模型是旧版 int8 还是 QDIT 量化模型。如果检测到旧版 int8 模型，会提示切换到 QDIT 以获得更好音质。QDIT 模型使用新的 ONNX 签名，具有改进的量化感知推理。
Vocoder Type Vocoder 类型
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Vocoder TypeVocoder 类型 | Default (vocoder_dml.onnx) or SiFiGAN (ICASSP 2023, optional, auto-fallback to default if model missing). Switching reloads only the vocoder session; main models stay loaded.Default（vocoder_dml.onnx）或 SiFiGAN（ICASSP 2023，可选，模型缺失时自动回退到默认）。切换仅重载 vocoder session，主模型保持已加载状态。 |
DefaultDefault |
| SiFiGAN PrecisionSiFiGAN 精度 | FP32 (recommended, full quality) or FP16 (low quality, cos≈0.95). Only visible when vocoder type is SiFiGAN. Missing variant auto-falls-back to the other.FP32（推荐，全质量）或 FP16（低质量，cos≈0.95）。仅在 vocoder 类型为 SiFiGAN 时可见。所选变体缺失时自动回退到另一变体。 | FP32FP32 |
DirectML VRAM Release Options DirectML 显存释放选项
These options only affect the DirectML (ORTNODE) backend.
这些选项仅对 DirectML（ORTNODE）后端生效。
| Option选项 | Description说明 | Default默认 |
|---|---|---|
| Release DML VRAM After Synthesis合成后释放 DirectML 显存 | Reloads diffStep and vocoder after each synthesis, forcing DirectML to reclaim intermediate tensor memory pools. Mitigates continuous-synthesis OOM at the cost of slightly longer post-synthesis wait.每次合成后重载 diffStep 和 vocoder，强制 DirectML 回收中间张量内存池。缓解连续合成 OOM，代价是合成后等待时间略增。 | Off关闭 |
| Release diffStep Before VocoderVocoder 推理前释放 diffStep | Releases the diffStep session after diffusion completes so the vocoder can use the freed ~3–4GB VRAM. Avoids 0x887A0006/TDR black screens. Cost: each vocoder run requires reloading diffStep (~1–3s), slowing multi-segment synthesis. Strongly recommended for SiFiGAN / low-VRAM GPUs.扩散完成后释放 diffStep session，让 vocoder 可使用释放的约 3–4GB 显存。避免 0x887A0006/TDR 黑屏。代价：每次 vocoder 后需重载 diffStep（~1–3 秒），多段合成会变慢。SiFiGAN / 低显存显卡强烈建议开启。 | On开启 |
Open Model Download 打开模型下载
Click Open Model Download to open the model download window for downloading or updating model files for the selected precision. Downloads use chunked parallel transfer (up to 16 concurrent connections) from ModelScope, with resume support (completed files are skipped on retry).
点击 Open Model Download 打开模型下载窗口，以下载或更新所选精度的模型文件。下载从 ModelScope 分片并行传输（最多 16 个并发连接），支持断点续传（重试时已完成文件会跳过）。
Resource Manager 资源管理器
Open from the menu bar: Settings > Resource Manager. It shows detected GPU devices (name, type, VRAM used/total) and lists all loaded models with options to Load All / Unload All or load/unload individual model groups. Models are automatically loaded when needed for synthesis; unloading frees VRAM for other applications.
从菜单栏打开：Settings > Resource Manager。它显示检测到的 GPU 设备（名称、类型、已用/总显存），并列出所有已加载模型，可 Load All / Unload All 或单独加载/卸载模型组。合成需要时模型会自动加载；卸载可释放显存供其他应用使用。