v1.12.0-alpha2

v1.12.0-alpha2

自 v1.11.0 正式版以来的全部改动汇总（含 1.12.0-alpha 预览版内容）。

• 实验性：同命令列表原地内联执行架构（NrInline 模式）：
  • 为《红色沙漠》（Crimson Desert）等当前无法进行帧切分的游戏特别修复，待实机测试；内联模式帧数待优化，请尝试内联菜单内多处实验性开关比较使用。
• Mochizuki 官方 a9b13dd 深度集成与 ACO 编译器支持：
  • 修复 ACO 解码器尾部（C64 到 fused-C32）屏障依赖与生命周期边界（感谢 @Overclockers）。
  • 支持可选 30 层降噪模式与 SR 后 NR 运动矢量直通。
• 《极限竞速：地平线 6》（Forza Horizon 6）兼容（感谢 B 站 清明yyyyy 测试反馈）：
  • 保留录制代理并规范增强屏障输入（感谢 @Overclockers，PR #8）。
  • 进程名含 forza 时自动开启「早期包装所有模块」（LmxxfEarlyAnyModuleWrap，需重启生效）与「允许增强屏障」（LmxxfAllowEnhancedBarriers，即时生效），二者均可在菜单强制开/关；地平线不使用内联模式。
• 《最终幻想 VII 重生》画面修复（感谢 B 站 清明yyyyy 测试反馈）：游戏中途切换颜色格式时自动转换，避免画面异常；格式不变的游戏不受影响。
• Mochizuki 移植本端 lmxxf 模式的曝光与白点（感谢 @Ciaranoff，PR #9）：
  • 新增「白点来源」下拉：关闭 / 跟随游戏 / 自行测光 / 自动，默认关闭（INI MochizukiWhitePointSource）。
  • 「白点/微调」滑块与之并列显示在 Mochizuki 输入页；选择自动来源时它是在自动结果上的亮度微调，1.0 为不调整。
• Mochizuki 其他改动：ACO 快速模式默认开启（已有 INI 中的 MochizukiAcoMode=false 不会被迁移）；安装时只提取权重；内联模式支持 FP16 输入，符合条件的内联链在计算队列上运行。
• 菜单：Mochizuki 预处理开启且为多层模式时，在人物保护处提示其不会生效；隐藏 AMD 后端不使用的 NGX 全局白点设置；标题栏与启动画面显示免费软件声明。
• 审查修复与发布加固：补入 LICENSE / NOTICE 与 lmxxf MIT 许可；可写 ACO 驱动缓存迁至 LOCALAPPDATA；内联中止的完成确认与回收；发布正文改由版本化文件生成。
• 新增若干默认关闭的内联诊断与实验开关（NrFrameDiag、NrGpuHandoff、NrInlineReuse、NrCaptureBackoff）。
• 架构稳定性与自动化测试回归。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置（已自带人物模型保留规则）。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：请从合法持有且许可允许使用的模型文件自行准备 native-game-tiled-assets 权重目录，提取方法见 原仓库。
• Mochizuki 0.0.4（引入 0.04/a9b13dd 特性）：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 本包已内置 PP-HumanSeg 与 YOLO11n-seg 人物分割模型权重（见 Licenses/）；不含 NVIDIA DLL、Daniel 安装器或全图专有神经渲染权重。

已知问题

• 实验性内联模式（NrInline）目前需在配置中显式开启，专用于解决不兼容命令列表拆分的游戏；在已兼容切分模式的游戏中建议保持默认。有限自旋预算不保证跨队列进度或帧时间上界。
• SR → NR 支持 DX12 及 DX11/Vulkan 转 DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• lmxxf 时序历史（抗闪烁）功能与 ViT 自适应复用继续互斥；建议关闭预测第 3 层。
• 部分场景与 XeFG 插件冲突，请关闭 XeFG 后重新进入游戏。

感谢各位用户的测试与反馈。

---

v1.12.0-alpha2

Summary of all changes since the v1.11.0 stable release (including the 1.12.0-alpha preview).

• Experimental: Same-Command-List Inline NR Execution (NrInline Mode):
  • A dedicated fix for games that currently cannot use frame splitting, such as Crimson Desert; pending in-game testing. Inline frame rate still needs optimization; please try the experimental switches in the Inline menu and compare.
• Mochizuki Official a9b13dd Upstream Integration & ACO Compiler Support:
  • Fixed ACO decoder-tail (C64 to fused-C32) barrier dependencies and lifetime boundaries (special thanks to @Overclockers).
  • Added selectable 30-layer pass mode and post-SR NR motion vector passthrough.
• Forza Horizon 6 compatibility (thanks to 清明yyyyy on Bilibili for testing and feedback):
  • Recording proxies are preserved and enhanced-barrier inputs normalized (thanks to @Overclockers, PR #8).
  • For executables whose name contains forza, early wrap of all modules (LmxxfEarlyAnyModuleWrap, restart required) and Allow enhanced barriers (LmxxfAllowEnhancedBarriers, applies live) turn on automatically; both can be forced on/off in the menu. Forza does not use inline mode.
• Final Fantasy VII Rebirth image fix (thanks to 清明yyyyy on Bilibili for testing and feedback): when the game switches colour format mid-session, NR now converts it automatically to avoid a broken image; games that keep one format are unaffected.
• Mochizuki now has the lmxxf-mode exposure and white point (thanks to @Ciaranoff, PR #9):
  • New "White point source" dropdown: Off / Follow game / Own meter / Auto, default Off (INI MochizukiWhitePointSource).
  • The "White point / brightness trim" slider now sits next to it on the Mochizuki input page; with an auto source it is a brightness trim on the automatic result, 1.0 = no change.
• Other Mochizuki changes: ACO fast mode is on by default (existing INIs with MochizukiAcoMode=false are not migrated); installation extracts only the weights; inline mode accepts FP16 input and eligible inline chains run on the compute queue.
• Menu: a warning under Person protection when Mochizuki preprocessing is on in multi-pass mode (protection does not apply); NGX-only global white point settings, unused by AMD backends, are hidden; the title bar and splash show a free-software notice.
• Review fixes and release hardening: LICENSE / NOTICE and the lmxxf MIT licence are included; the writable ACO driver cache moved to LOCALAPPDATA; inline aborts are certified and reclaimed; the release body is generated from a versioned notes file.
• Several default-off inline diagnostics and experimental switches (NrFrameDiag, NrGpuHandoff, NrInlineReuse, NrCaptureBackoff).
• Architectural Stability & Automated Test Coverage.

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run Setup.bat. When updating, we recommend choosing Y to uninstall first and start with default settings (existing person models are preserved automatically).
• Daniel 0.6.0: obtain runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: prepare native-game-tiled-assets from model files you lawfully possess and are permitted to use; extraction instructions are in the upstream repository.
• Mochizuki 0.0.4 (includes 0.04/a9b13dd features): extract weights from nvngx_dlssnr.dll 310.8.0 (same file as Daniel). Python 3.10+ is required.
• Bundles PP-HumanSeg and YOLO11n-seg person partition models (see Licenses/); NVIDIA DLLs, Daniel installers, and proprietary neural weights are not included.

Known issues

• Experimental NrInline mode must be explicitly enabled in configuration; intended specifically for games with split-ineligible command lists. Keep default split mode for normally compatible titles. Bounded spinning does not guarantee cross-queue progress or a frame-time limit.
• SR → NR supports DX12 and DX11/Vulkan-to-DX12 bridges; does not provide native RR/Vulkan. Unsupported layouts retain the SR frame and display diagnostic reasons.
• lmxxf temporal history and adaptive ViT reuse remain mutually exclusive. Disabling predicted pass 3 is recommended.
• May conflict with the XeFG plugin in certain titles; disable XeFG and restart the game if encountered.

Thanks to all users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：c1cb8c06389af7fc2580757ce00cc6efd416f111

What's Changed

• fix(nr): 修复代理传递及混合屏障输入准入 by @Overclockers in #8
• feat(nr): mochizuki meters the frame's own exposure when the game supplies none by @Ciaranoff in #9

New Contributors

• @Ciaranoff made their first contribution in #9

Full Changelog: v1.12.0-alpha...v1.12.0-alpha2

v1.12.0-alpha

v1.12.0-alpha

• 实验性：同命令列表原地内联执行架构（NrInline 模式）：
  • 解决《红色沙漠》（Crimson Desert）等游戏中因跨 Evaluate 未闭合非时间戳 Query（open_query）或复杂增强屏障导致的切分拒绝与崩溃。
• Mochizuki 0.0.4 官方 a9b13dd 深度集成与 ACO 编译器支持：
  • 修复 ACO 解码器尾部（C64 到 fused-C32）屏障依赖与生命周期边界（感谢 @Overclockers）。
  • 支持可选 30 层降噪模式与 SR 后 NR 运动矢量直通。
• 架构稳定性与自动化测试回归。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置（已自带人物模型保留规则）。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.4：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 本包已内置 PP-HumanSeg 与 YOLO11n-seg 人物分割模型权重（见 Licenses/）；不含 NVIDIA DLL、Daniel 安装器或全图专有神经渲染权重。

已知问题

• 实验性内联模式（NrInline）目前需在配置中显式开启，专用于解决不兼容命令列表拆分的游戏；在已兼容切分模式的游戏中建议保持默认。
• SR → NR 支持 DX12 及 DX11/Vulkan 转 DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• lmxxf 时序历史（抗闪烁）功能与 ViT 自适应复用继续互斥；建议关闭预测第 3 层。
• 部分场景与 XeFG 插件冲突，请关闭 XeFG 后重新进入游戏。

感谢各位用户的测试与反馈。

---

v1.12.0-alpha

• Experimental: Same-Command-List Inline NR Execution (NrInline Mode):
  • Resolves split rejection (open_query) and crashes in games like Crimson Desert caused by unclosed non-timestamp queries or enhanced barriers across Evaluate.
• Mochizuki Official a9b13dd Upstream Integration & ACO Compiler Support:
  • Fixed ACO decoder-tail (C64 to fused-C32) barrier dependencies and lifetime boundaries (special thanks to @Overclockers).
  • Added selectable 30-layer pass mode and post-SR NR motion vector passthrough.
• Architectural Stability & Automated Test Coverage.

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run Setup.bat. When updating, we recommend choosing Y to uninstall first and start with default settings (existing person models are preserved automatically).
• Daniel 0.6.0: obtain runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or mirror.
• Mochizuki 0.0.4 (includes 0.04/a9b13dd features): extract weights from nvngx_dlssnr.dll 310.8.0 (same file as Daniel). Python 3.10+ is required.
• Bundles PP-HumanSeg and YOLO11n-seg person partition models (see Licenses/); NVIDIA DLLs, Daniel installers, and proprietary neural weights are not included.

Known issues

• Experimental NrInline mode must be explicitly enabled in configuration; intended specifically for games with split-ineligible command lists. Keep default split mode for normally compatible titles.
• SR → NR supports DX12 and DX11/Vulkan-to-DX12 bridges; does not provide native RR/Vulkan. Unsupported layouts retain the SR frame and display diagnostic reasons.
• lmxxf temporal history and adaptive ViT reuse remain mutually exclusive. Disabling predicted pass 3 is recommended.
• May conflict with the XeFG plugin in certain titles; disable XeFG and restart the game if encountered.

Thanks to all users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：cc780a212b1d1748f509113fb3dbed750914adb6

Full Changelog: v1.11.0...v1.12.0-alpha

v1.11.0

1.11.0（相较于 1.10.5-alpha）

• 定向整合上游 OptiScaler UI 升级、稳定性与字体排版修复：
  • 采用响应式卡片分类布局，菜单在各类窗口及分辨率下自适应排版。
  • 保留三后端独立控制、分组 reset 与中英双语本地化支持，增加可选背景模糊（MenuBlur）。
  • 页面切换渲染前主动刷新活动后端状态，彻底解决换页后状态滞后。
  • 修复零尺寸 HWND 交换链包装问题，修复窗口调整大小时的潜在崩溃；修复 FFX 超分辨率版本查询在空列表或失败情况下的空指针问题。
  • 重构并扩充 NotoSansSC-Menu.ttf 中文字体子集，增补 Mochizuki 专属分项耗时（融/拆/张/浮）与人物保护全部字符，实测 144 种界面布局 0 乱码/豆腐块。
• 增加 Authenticode 自签名：
  • 签署 staging 二进制并生成 SIGNATURES.json 强绑定清单，提升兼容性与防篡改验证。
• 实验性：公共残差分频与细节/肤色保护：
  • 新增低频增益、细节增益及肤色/边缘保护启发式控制（NrResidualLowGain、NrResidualDetailGain、NrResidualSkinProtection、NrResidualEdgeProtection）。
  • 适用于三后端统一输出，在保持原始色域与 alpha 基础上增强画面立体感与边缘细节。
• 实验性：人物/场景分区分割、双后端通用保护与时序抗鬼影（支持 Mochizuki 与 lmxxf 后端，当前仅使用 Mochizuki 在《燕云十六声》游戏中做了多轮测试）：
  • 引入基于 ONNX Runtime CPU 的异步轻量推理通道（NrPersonPartition）。
  • 通用打通 lmxxf 与 Mochizuki 双后端的人物保护控制（NrPersonStrength 强度 / NrPersonDetailGain 细节保留）。
  • 加入运动对齐的蒙版时序平滑、过期淡出与遮挡拒绝，减轻蒙版刷新时的跳变；快速运动效果仍需游戏验证。
• 开箱即用体验与安装器自愈能力增强：
  • 安装包正式直接内置 ONNX Runtime 动态库 (MIT) 以及 PP-HumanSeg (Apache-2.0) 与 YOLO11n-seg (AGPL-3.0) 权重，默认启用 PP-HumanSeg（支持在菜单内自由切换），解压即可直接开箱启用人物保护，支持自备模型无缝覆盖。
  • Setup.bat 重装/升级防丢：更新时选择 Y 先卸载 也会智能保留并继承已存在的 person-model，避免重复下载与意外清空。
  • 人物推理使用独立 person-worker.exe，ONNX 与新版 VC++ 运行库放在 person-model/；使用普通优先级与 2–4 个 CPU 线程。
• Mochizuki 后端与 ReShade 设备穿透兼容：
  • 修复 ReShade 包装设备与原生 Fence 设备身份不同导致 Mochizuki 提交被拒绝的问题，已在燕云组合测试通过；其他加载顺序及插件仍需验证。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置（已自带人物模型保留规则）。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3（引入0.04部分特性）：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包已内置 PP-HumanSeg 与 YOLO11n-seg 人物分割模型权重（见 Licenses/）；不含 NVIDIA DLL、Daniel 安装器或全图专有神经渲染权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan 转 DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• lmxxf 时序历史（抗闪烁）功能与 ViT 自适应复用继续互斥；SR 后 NR 路径暂不启用 lmxxf 时序历史。预测第 3 层使用最后一次真实网络的辅助量，画面仍可能存在近似误差，建议关闭预测第 3 层。
• ReShade 兼容性：已修复所测燕云组合的 Mochizuki 提交问题；如其他游戏或注入组合仍无法生效，请附日志反馈。
• 部分游戏如遇 NR 丢失带来闪烁或 hook 失效问题，请尝试开启实验性开关“提前接管”（Early command-list wrap / INI: LmxxfEarlyExeWrap，设为 force on，伊莫已默认开启）。
• 部分场景与 XeFG 插件冲突，请关闭 XeFG 后重新进入游戏。

感谢各位 Bilibili 用户的测试与反馈。

---

1.11.0 (compared to 1.10.5-alpha)

• Selective Upstream OptiScaler UI Upgrade, Stability and Chinese Typographic Fixes:
  • Responsive card categorization layout adapting across window sizes and resolutions.
  • Retains dedicated controls for all three backends, group reset, bilingual localization, and optional background blur (MenuBlur).
  • Actively refreshes active backend state before rendering any page to prevent UI display lag.
  • Fixed zero-sized HWND swapchain wrapping to prevent window resize crashes; rejected failed or empty upscaler version queries in FFX discovery.
  • Rebuilt and expanded Chinese font subset (NotoSansSC-Menu.ttf) with Mochizuki timing metrics glyphs (融/拆/张/浮) and person protection terminology for complete 100% coverage with zero tofu characters.
• Added Authenticode Self-Signing:
  • Signs staged binaries and records bound hash manifests in SIGNATURES.json for tamper-resistance and loader compatibility.
• Experimental: Shared Residual Shaping and Detail/Skin Protection:
  • Added low/detail gain and skin/edge protection heuristics (NrResidualLowGain, NrResidualDetailGain, NrResidualSkinProtection, NrResidualEdgeProtection).
  • Applied across all backends' common outputs while keeping original color space and alpha intact.
• Experimental: Person/Scene Partition, Shared Detail Controls & Anti-Ghosting (supports Mochizuki and lmxxf backends; currently multi-round tested only on Mochizuki in Where Winds Meet):
  • Asynchronous CPU inference path via ONNX Runtime supporting selectable PP-HumanSeg (default) and YOLO11n-seg models (NrPersonPartition).
  • Shared person protection controls across lmxxf and Mochizuki (NrPersonStrength and NrPersonDetailGain).
  • Added motion-aligned temporal mask smoothing, expiry fading and disocclusion rejection to reduce mask refresh pulses. Fast-motion results still need game validation.
• Out-of-the-Box Packaging & Installer Self-Healing:
  • Bundles ONNX Runtime (MIT) along with PP-HumanSeg (Apache-2.0) and YOLO11n-seg (AGPL-3.0) weights under person-model/, defaulting to PP-HumanSeg (with in-menu switching) for out-of-the-box evaluation with custom model override support.
  • Setup.bat preserves existing person-model assets when upgrading with uninstallFirst (selecting Y to uninstall).
  • Runs person inference in a separate person-worker.exe with private ONNX/VC++ dependencies under person-model/. Uses normal priority and 2–4 CPU threads.
• Mochizuki & ReShade Native Fence Penetration:
  • Resolves D3D12 device wrapper mismatches behind ReShade by safely extracting native device identities through session fences (accept proven native fence identity behind ReShade).

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings (existing person models are preserved automatically).
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3 (includes partial 0.0.4 features): extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• Bundles PP-HumanSeg and YOLO11n-seg person partition models (see Licenses/); NVIDIA DLLs, the Daniel installer and full-frame neural rendering weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• lmxxf temporal history (anti-flicker) and adaptive ViT reuse remain mutually exclusive; lmxxf temporal history is unavailable in the post-SR NR path. Predicted pass 3 uses auxiliary data from the last real network execution and may introduce approximation artifacts; disabling predicted pass 3 is recommended.
• ReShade compatibility: the tested Where Winds Meet combination now submits Mochizuki frames correctly. Other games, load orders and add-ons still need validation; please include logs with reports.
• If NR dropping out causes flicker or hook failure in some games, try enabling the experimental "Early command-list wrap" switch (INI: LmxxfEarlyExeWrap, set to force on; enabled by default for Aniimo).
• May conflict with the XeFG plugin in certain scenarios; please disable XeFG and restart the game if encountered.

Thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：335aadee21b88ecdcc5b1a02ae80a36cf0722016

Full Changelog: v1.10.4.1...v1.11.0

v1.10.5-alpha

1.10.5-alpha

• 接入 mochi 0.0.4 部分特性（无重大更新）

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3（引入0.04部分特性）：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包不含 NVIDIA DLL、Daniel 安装器或模型权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan → DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• lmxxf时序历史（抗闪烁）功能与 ViT 自适应复用继续互斥；SR 后NR路径暂不启用 lmxxf 时序历史。预测第 3 层使用最后一次真实网络的辅助量，画面仍可能存在近似误差，建议关闭预测第 3 层。
• 可能与 ReShade 冲突（Mochizuki 后端已在 1.10.4.1 尝试兼容修复；若仍遇冲突请暂时移除 ReShade）。
• 部分游戏如遇 NR 丢失带来闪烁或 hook 失效问题，请尝试开启实验性开关“提前接管”（Early command-list wrap / INI: LmxxfEarlyExeWrap，设为 force on，伊莫已默认开启）。
• 部分场景与 XeFG 插件冲突，请关闭 XeFG 后重新进入游戏。

补充感谢

• 补充感谢 @Overclockers 在 1.10.3 版本贡献的 NR 激活时独立 XeFG 插帧倍率覆盖功能（PR #3）。注：仍需自备 XeFG 补丁，参考 README 对应“可选：3倍及以上多帧生成 (XeFG/DLSSG)”章节。
• 各位 Bilibili 用户的测试与反馈。

---

1.10.5-alpha

• Integrated partial features from Mochizuki v0.0.4 (no major changes)

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings.
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3 (includes partial 0.0.4 features): extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• NVIDIA DLLs, the Daniel installer and model weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• lmxxf temporal history (anti-flicker) and adaptive ViT reuse remain mutually exclusive; lmxxf temporal history is unavailable in the post-SR NR path. Predicted pass 3 uses auxiliary data from the last real network execution and may introduce approximation artifacts; disabling predicted pass 3 is recommended.
• May conflict with ReShade (attempted compatibility fix for Mochizuki in 1.10.4.1; remove ReShade temporarily if issues persist).
• If NR dropping out causes flicker or hook failure in some games, try enabling the experimental "Early command-list wrap" switch (INI: LmxxfEarlyExeWrap, set to force on; enabled by default for Aniimo).
• May conflict with the XeFG plugin in certain scenarios; please disable XeFG and restart the game if encountered.

Additional thanks

• Special thanks to @Overclockers for contributing the NR-only XeFG multiplier override in version 1.10.3 (PR #3). Note: XeFG unlock files are still required; see the "Optional: 3x+ Frame Generation (XeFG/DLSSG)" section in README for details.
• Thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：d637147e6e62383faa99edaa3f4e45bb5ed2347e

Full Changelog: v1.10.4.1...v1.10.5-alpha

v1.10.4.1

1.10.4.1（相比 1.10.3.1）

• 内置并支持安装 DLSS 伪装增强（OptiPatcher 原生功能）：安装结束时提示用户是否安装；支持自动获取官方最新脚本与组件并同步缓存，若超时（60s）或离线则自动回退至内置离线版本，并在 OptiScaler.ini 中自动开启 LoadAsiPlugins=true。
• lmxxf 时序历史（防闪烁）支持 SR 前的 1–3 层 NR：中间层只传递网络结果，最后一次真实网络导出时序辅助量，最终画面统一平滑。预测第 3 层仍只运行两次真实网络；切换层数时重置历史。
• 完整同步 lmxxf 官方 PR12 主线闭包：解除所有本地临时源码 Pin 与补丁覆盖，直接跟随官方最新合并提交，保留项目私有的时序历史算法与配置优先级。
• 定向同步 OptiScaler 上游的 DLSS/DLSSG 多副本生命周期：新副本接管失败不再撤销旧副本，返回回调按模块隔离，卸载后安全退休。
• 改进 DX11 → DX12 + XeFG 的交换链协调：确认参数等价时避免重复调整伴随交换链，正确处理窗口尺寸与 present 同步；GPU 等待失败时保留在用资源。
• 减少启用 HUD/描述符跟踪时批量复制的重复查堆开销；单个描述符保留原路径。
• 尝试修复 Mochizuki 后端与 ReShade 不兼容问题：恢复包装层背后的逻辑命令列表（感谢 @MatheusFerreiraS 的兼容思路参考）。
• 修复 AMD 驱动扩展 Hook（amdxc64）并发初始化竞态：安全串行化并安全发布 trampoline，解决部分游戏开局偶发性闪退崩溃问题。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包不含 NVIDIA DLL、Daniel 安装器或模型权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan → DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• lmxxf时序历史（抗闪烁）功能与 ViT 自适应复用继续互斥；SR 后NR路径暂不启用 lmxxf 时序历史。预测第 3 层使用最后一次真实网络的辅助量，画面仍可能存在近似误差，建议关闭预测第 3 层。
• 可能与 ReShade 冲突（Mochizuki 后端已在 1.10.4.1 尝试兼容修复；若仍遇冲突请暂时移除 ReShade）。
• 部分游戏如遇 NR 丢失带来闪烁或 hook 失效问题，请尝试开启实验性开关“提前接管”（Early command-list wrap / INI: LmxxfEarlyExeWrap，设为 force on，伊莫已默认开启）。
• 部分场景与 XeFG 插件冲突，请关闭 XeFG 后重新进入游戏。

补充感谢

• 补充感谢 @Overclockers 在 1.10.3 版本贡献的 NR 激活时独立 XeFG 插帧倍率覆盖功能（PR #3）。注：仍需自备 XeFG 补丁，参考 README 对应“可选：3倍及以上多帧生成 (XeFG/DLSSG)”章节。
• 各位 Bilibili 用户的测试与反馈。

---

1.10.4.1 (compared to 1.10.3.1)

• Integrated OptiPatcher (DLSS Spoofing Enhancement): Setup now prompts to install OptiPatcher, attempting to fetch the latest official rolling release with a 60-second timeout, falling back gracefully to the bundled offline version if offline, and automatically enabling LoadAsiPlugins=true.
• lmxxf temporal history (anti-flicker) now supports 1–3 NR passes before SR. Intermediate passes forward network results; only the final real network exports temporal auxiliary data, and smoothing applies to the final image. Predicted pass 3 still uses two real network executions. Changing the pass count resets history.
• Synchronized with official lmxxf PR12 mainline closure: Removed local temporary source pins and patches to follow the official merge, while preserving project-owned temporal history algorithms and configuration priority.
• Selectively integrated OptiScaler DLSS/DLSSG plugin lifetime improvements: a failed new hook preserves the older instance, callbacks remain module-specific, and unloaded instances retire safely.
• Improved swapchain coordination for DX11 → DX12 + XeFG: avoid redundant companion resize only when parameters are confirmed equivalent, handle window sizing and present synchronization, and preserve in-use resources when a GPU wait fails.
• Reduced repeated heap lookups in batched descriptor copies when HUD/descriptor tracking is enabled; single-descriptor copies retain their original path.
• Attempted compatibility fix for Mochizuki backend with ReShade: recovers logical command lists behind ReShade wrappers (special thanks to @MatheusFerreiraS for reference implementation).
• Fixed concurrency race condition in AMD driver extension hook (amdxc64): serialized initialization and safely published trampolines, resolving intermittent startup crash in some games.

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings.
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3: extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• NVIDIA DLLs, the Daniel installer and model weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• lmxxf temporal history (anti-flicker) and adaptive ViT reuse remain mutually exclusive; lmxxf temporal history is unavailable in the post-SR NR path. Predicted pass 3 uses auxiliary data from the last real network execution and may introduce approximation artifacts; disabling predicted pass 3 is recommended.
• May conflict with ReShade (attempted compatibility fix for Mochizuki in 1.10.4.1; remove ReShade temporarily if issues persist).
• If NR dropping out causes flicker or hook failure in some games, try enabling the experimental "Early command-list wrap" switch (INI: LmxxfEarlyExeWrap, set to force on; enabled by default for Aniimo).
• May conflict with the XeFG plugin in certain scenarios; please disable XeFG and restart the game if encountered.

Additional thanks

• Special thanks to @Overclockers for contributing the NR-only XeFG multiplier override in version 1.10.3 (PR #3). Note: XeFG unlock files are still required; see the "Optional: 3x+ Frame Generation (XeFG/DLSSG)" section in README for details.
• Thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：80ba33ff209b16096dcf9ca4067040b169276aa8

Full Changelog: v1.10.4...v1.10.4.1

v1.10.4

1.10.4（相比 1.10.3.1）

• 内置并支持安装 DLSS 伪装增强（OptiPatcher 原生功能）：安装结束时提示用户是否安装；支持自动获取官方最新脚本与组件并同步缓存，若超时（60s）或离线则自动回退至内置离线版本，并在 OptiScaler.ini 中自动开启 LoadAsiPlugins=true。
• lmxxf 时序历史（防闪烁）支持 SR 前的 1–3 层 NR：中间层只传递网络结果，最后一次真实网络导出时序辅助量，最终画面统一平滑。预测第 3 层仍只运行两次真实网络；切换层数时重置历史。
• 完整同步 lmxxf 官方 PR12 主线闭包：解除所有本地临时源码 Pin 与补丁覆盖，直接跟随官方最新合并提交，保留项目私有的时序历史算法与配置优先级。
• 定向同步 OptiScaler 上游的 DLSS/DLSSG 多副本生命周期：新副本接管失败不再撤销旧副本，返回回调按模块隔离，卸载后安全退休。
• 改进 DX11 → DX12 + XeFG 的交换链协调：确认参数等价时避免重复调整伴随交换链，正确处理窗口尺寸与 present 同步；GPU 等待失败时保留在用资源。
• 减少启用 HUD/描述符跟踪时批量复制的重复查堆开销；单个描述符保留原路径。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包不含 NVIDIA DLL、Daniel 安装器或模型权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan → DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• lmxxf时序历史（抗闪烁）功能与 ViT 自适应复用继续互斥；SR 后NR路径暂不启用 lmxxf 时序历史。预测第 3 层使用最后一次真实网络的辅助量，画面仍可能存在近似误差，建议关闭预测第 3 层。
• 可能与 ReShade 冲突（已知Mochizuki）。
• 部分游戏如遇 NR 丢失带来闪烁或 hook 失效问题，请尝试开启实验性开关“提前接管”（Early command-list wrap / INI: LmxxfEarlyExeWrap，设为 force on，伊莫已默认开启）。
• 部分场景与 XeFG 插件冲突，请关闭 XeFG 后重新进入游戏。

补充感谢

• 补充感谢 @Overclockers 在 1.10.3 版本贡献的 NR 激活时独立 XeFG 插帧倍率覆盖功能（PR #3）。注：仍需自备 XeFG 补丁，参考 README 对应“可选：3倍及以上多帧生成 (XeFG/DLSSG)”章节。
• 各位 Bilibili 用户的测试与反馈。

---

1.10.4 (compared to 1.10.3.1)

• Integrated OptiPatcher (DLSS Spoofing Enhancement): Setup now prompts to install OptiPatcher, attempting to fetch the latest official rolling release with a 60-second timeout, falling back gracefully to the bundled offline version if offline, and automatically enabling LoadAsiPlugins=true.
• lmxxf temporal history (anti-flicker) now supports 1–3 NR passes before SR. Intermediate passes forward network results; only the final real network exports temporal auxiliary data, and smoothing applies to the final image. Predicted pass 3 still uses two real network executions. Changing the pass count resets history.
• Synchronized with official lmxxf PR12 mainline closure: Removed local temporary source pins and patches to follow the official merge, while preserving project-owned temporal history algorithms and configuration priority.
• Selectively integrated OptiScaler DLSS/DLSSG plugin lifetime improvements: a failed new hook preserves the older instance, callbacks remain module-specific, and unloaded instances retire safely.
• Improved swapchain coordination for DX11 → DX12 + XeFG: avoid redundant companion resize only when parameters are confirmed equivalent, handle window sizing and present synchronization, and preserve in-use resources when a GPU wait fails.
• Reduced repeated heap lookups in batched descriptor copies when HUD/descriptor tracking is enabled; single-descriptor copies retain their original path.

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings.
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3: extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• NVIDIA DLLs, the Daniel installer and model weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• lmxxf temporal history (anti-flicker) and adaptive ViT reuse remain mutually exclusive; lmxxf temporal history is unavailable in the post-SR NR path. Predicted pass 3 uses auxiliary data from the last real network execution and may introduce approximation artifacts; disabling predicted pass 3 is recommended.
• May conflict with ReShade (confirmed with Mochizuki).
• If NR dropping out causes flicker or hook failure in some games, try enabling the experimental "Early command-list wrap" switch (INI: LmxxfEarlyExeWrap, set to force on; enabled by default for Aniimo).
• May conflict with the XeFG plugin in certain scenarios; please disable XeFG and restart the game if encountered.

Additional thanks

• Special thanks to @Overclockers for contributing the NR-only XeFG multiplier override in version 1.10.3 (PR #3). Note: XeFG unlock files are still required; see the "Optional: 3x+ Frame Generation (XeFG/DLSSG)" section in README for details.
• Thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：1f2b6dcdb644c6dcb39b4628c7321e7fc689abc6

Full Changelog: v1.10.3.1...v1.10.4

v1.10.3.1

1.10.3.1（相比 1.10.3）

• 修复 Page Up 性能叠加层（Performance Overlay / FPS Overlay） 布局折叠异常
• 默认字体大小由 14.0 调整为 16.0
• 性能叠加层新增**独立字体调色盘（Text Color）**与重置按钮（Reset）
• 勾选“跟随全局主题色（Use Theme Colors）”时，独立调色盘自动置灰禁用并提供悬浮说明

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包不含 NVIDIA DLL、Daniel 安装器或模型权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan → DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• 时序历史与 ViT 自适应复用的兼容性仍待调查；多层或 SR 后路径暂不启用 lmxxf 时序历史。部分场景仍可能闪烁，可尝试 Mochizuki。
• 可能与 ReShade 冲突。
• Unity 游戏如遇 NR 丢失带来的闪烁，请尝试开启对应实验性开关（伊莫已默认开启）。

特别感谢：各位 Bilibili 用户的测试与反馈。

---

1.10.3.1 (compared to 1.10.3)

• Fixed layout collapse in the Page Up Performance Overlay / FPS Overlay
• Updated default font size from 14.0 to 16.0
• Added an independent Text Color picker and Reset button for the Performance Overlay
• Grayed out the overlay color picker with a tooltip when "Use Theme Colors" is enabled

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings.
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3: extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• NVIDIA DLLs, the Daniel installer and model weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• History/adaptive-ViT compatibility remains under investigation. lmxxf temporal history is unavailable with multiple passes or after SR. Some scenarios may still flicker; try Mochizuki if affected.
• ReShade may conflict.
• If NR dropping out causes flicker in a Unity game, try enabling the corresponding experimental option. It is enabled by default for Aniimo.

Special thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：76dc9fe2a548a8052ab44458f636a7ba3a0c3d59

Full Changelog: v1.10.3...v1.10.3.1

v1.10.3

1.10.3（相比 1.10.2）

• Ins 菜单新增 English / 简体中文，默认英文；语言选框位于“保存设置”旁，保存后下次启动继续使用。
• 默认菜单统一使用内嵌 Noto Sans SC，改善中文、英文和数字混排时的大小与风格差异，无需安装字体或中文语言包；保留自定义字体设置。
• 精简中文标签，补充选项说明，调整菜单布局与重置范围。三后端的详细状态集中到“诊断”，主界面保留简短状态、可用的网络耗时及失败／回退提示。
• 新增 lmxxf Temporal history (anti-flicker) / 时序历史（防闪烁），默认开启（已有 INI 选择优先）；目前支持 SR 前的单层 NR，需要有效运动矢量与深度。可减轻部分闪烁，也会增加 GPU 耗时，可能带来拖影。
• 开启时序历史时暂时禁用 ViT adaptive reuse 及四项参数，避免组合启用导致严重闪烁；保留原偏好，关闭历史后恢复。减少历史处理中的拷贝和重复写入。
• 修复 SR → NR 的处理尺寸错误与保留录制间临时资源复用造成的闪烁，修复原生 4K 时序历史的调度边界。
• 增加 NR 专用 XeFG 倍率设置；修正后端初始化期间 Vulkan 调用保护的线程隔离。
• 定向跟进 OptiScaler 97e99b4c 的 NVAPI、DXGI SwapChain 和 HUD 捕获稳定性修复，保留三后端生命周期；并非全量合并。lmxxf 仍为 0.41。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包不含 NVIDIA DLL、Daniel 安装器或模型权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan → DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• 时序历史与 ViT 自适应复用的兼容性仍待调查；多层或 SR 后路径暂不启用 lmxxf 时序历史。部分场景仍可能闪烁，可尝试 Mochizuki。
• 可能与 ReShade 冲突。
• Unity 游戏如遇 NR 丢失带来的闪烁，请尝试开启对应实验性开关（伊莫已默认开启）。

特别感谢：各位 Bilibili 用户的测试与反馈。

---

1.10.3 (compared to 1.10.2)

• Added English / Simplified Chinese to the Ins menu, with English as the default. The language selector sits beside Save Settings; saving retains the choice for the next launch.
• The default menu now uses embedded Noto Sans SC for Chinese, English and numerals, improving consistency in mixed text. No font or Chinese language-pack installation is required; custom font settings are preserved.
• Shortened Chinese labels, expanded help text, and refined menu layout and reset scopes. Detailed status for all three backends is grouped under Diagnostics; compact status, available network timings, and failure/fallback reasons remain visible in the main panel.
• Added lmxxf Temporal history (anti-flicker), enabled by default while preserving existing INI preferences. It currently supports one NR pass before SR, with valid motion/depth guides. It may reduce flicker, at additional GPU cost and with possible ghosting.
• Temporal history temporarily disables ViT adaptive reuse and its four controls to avoid severe flicker when combined. Saved preferences return when history is disabled. Reduced copies and redundant writes in history processing.
• Fixed SR → NR processing dimensions and flicker from scratch-resource reuse across retained recordings; fixed native 4K history dispatch bounds.
• Added an NR-only XeFG multiplier setting and isolated Vulkan call guards per thread during backend initialization.
• Selectively ported NVAPI, DXGI swapchain and HUD capture stability fixes reviewed against OptiScaler 97e99b4c, preserving the three-backend lifecycle. This is not a full upstream merge. lmxxf remains at 0.41.

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings.
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3: extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• NVIDIA DLLs, the Daniel installer and model weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• History/adaptive-ViT compatibility remains under investigation. lmxxf temporal history is unavailable with multiple passes or after SR. Some scenarios may still flicker; try Mochizuki if affected.
• ReShade may conflict.
• If NR dropping out causes flicker in a Unity game, try enabling the corresponding experimental option. It is enabled by default for Aniimo.

Special thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：3bbfff9c652acbf4c3868cef5d100178aa9e79b1

What's Changed

• Add NR-only XeFG multiplier override by @Overclockers in #3

New Contributors

• @Overclockers made their first contribution in #3

Full Changelog: v1.10.2...v1.10.3

v1.10.2

1.10.2（相比 1.10.1）

• 新增实验性 SR → NR：在超分后执行神经渲染。
• lmxxf 更新到 0.41，新增肤色保护、第 3 层预测（尽可能还原 3 层 NR 效果，同时保留接近 2 层的帧数）。选择 3 层时，默认执行两次真实网络＋第三层局部预测；关闭 Predict third pass (lossy) 后执行三次真实网络。肤色保护默认关闭，自由分辨率默认开启。
• 改进 Unity 部分场景因丢失 NR 造成的闪烁：默认覆盖《伊莫》（Aniimo）的 UnityPlayer 调用；其他 Unity 游戏可在 Early command-list wrap 中手动启用，保存后重启生效。
• 修复旧 lmxxf 着色器残留遮蔽新包、导致彩色噪点或条块的问题。
• 修复 Mochizuki 遇到短暂分辨率跳变时造成的闪烁。
• 改进安装与卸载逻辑。
• 调整 Ins 菜单与 INI 文件备注。

安装与对应后端权重准备

• 请完整解压 ZIP 至任意文件夹并运行新包的 Setup.bat。更新时推荐选择 Y 先卸载，重新使用包内默认配置。
• Daniel 0.6.0：运行文件请前往 Daniel 官方发布页 获取；nvngx_dlssnr.dll 310.8.0 请自行获取。
• lmxxf 0.41：需自备 native-game-tiled-assets 权重目录，见 原仓库 或 下载镜像。
• Mochizuki 0.0.3：从 nvngx_dlssnr.dll 310.8.0（与 Daniel 使用同一文件）提取权重，需要 Python 3.10+。
• 如本次安装目标游戏文件夹内留存曾安装过的三后端有效权重，执行 Setup.bat 后可自动利用，并补回安装包内缺失的权重，使得日常更新本插件时无需反复重新拷贝。
• 本包不含 NVIDIA DLL、Daniel 安装器或模型权重。

已知问题

• SR → NR 支持 DX12 及 DX11/Vulkan → DX12 桥接，不扩展原生 RR/Vulkan；不支持的纹理布局保留 SR 图像并显示原因。
• 部分游戏场景的 lmxxf 闪烁问题仍在修复，遇到闪烁时推荐尝试 Mochizuki 作为后端。
• 可能与 ReShade 冲突。
• Unity 游戏如遇 NR 丢失带来的闪烁，请尝试开启对应实验性开关（伊莫已默认开启）。

特别感谢：各位 Bilibili 用户的测试与反馈。

---

1.10.2 (compared to 1.10.1)

• Added experimental SR → NR: run neural rendering after upscaling.
• Updated lmxxf to 0.41, adding skin protection and third-pass prediction, designed to approximate three-pass NR quality while retaining frame rates close to two passes. With three passes selected, the default is two real network passes plus a local third-pass prediction. Disable Predict third pass (lossy) to run three real network passes. Skin protection defaults to off; native free resolution defaults to on.
• Improved flicker caused by NR dropping out in some Unity scenarios. Aniimo's UnityPlayer calls are covered by default; other Unity games can enable Early command-list wrap manually. Save settings and restart the game to apply.
• Fixed stale lmxxf shaders overriding an updated package and causing colored noise or blocks.
• Fixed Mochizuki flicker caused by brief resolution changes.
• Improved installation and uninstallation logic.
• Refined the Ins menu and INI comments.

Installation and backend model preparation

• Extract the complete ZIP into a folder of your choice and run the new package's Setup.bat. When upgrading, we recommend choosing Y to uninstall first and start with the package's default settings.
• Daniel 0.6.0: obtain the runtime files from the Daniel releases page. Obtain nvngx_dlssnr.dll 310.8.0 separately.
• lmxxf 0.41: provide native-game-tiled-assets weights from the upstream repository or download mirror.
• Mochizuki 0.0.3: extract the weights from nvngx_dlssnr.dll 310.8.0, the same file used by Daniel. Python 3.10+ is required.
• If the game folder selected for this installation still contains valid weights from previous installations of any of the three backends, Setup.bat can reuse them and copy missing weights back into the installer folder, so you do not need to copy them manually again for routine plugin updates.
• NVIDIA DLLs, the Daniel installer and model weights are not included.

Known issues

• SR → NR supports DX12 and the DX11/Vulkan-to-DX12 bridges; it does not add native RR/Vulkan support. Unsupported texture layouts preserve the SR image and report the reason.
• lmxxf flicker in some game scenarios is still being addressed. If you encounter flicker, we recommend trying Mochizuki as the backend.
• ReShade may conflict.
• If NR dropping out causes flicker in a Unity game, try enabling the corresponding experimental option. It is enabled by default for Aniimo.

Special thanks to all Bilibili users for testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Source / 源码：e62f52d84418b8e8c55eab389a8529da9e4851b8

Full Changelog: v1.10.1...v1.10.2

v1.10.1

1.10.1（相比 1.10.0）

• 修复 Mochizuki 后端在 AMD 26.9.2+ 新版显卡驱动中无法正常使用（着色器编译失败）的问题，感谢 @MatheusFerreiraS 提供的修复方案。

Daniel 后端

运行文件、权重以及 nvngx_dlssnr.dll 310.8.0 请自行前往 Daniel 官方发布页 下载。本版继续支持 Daniel 0.6.0。

lmxxf 后端

权重文件：lmxxf 原仓库 或 网盘下载镜像。

Mochizuki 后端

提取模型需要自行提供受支持的 nvngx_dlssnr.dll 310.8.0（同 Daniel 后端）以及 Python 3.10+，或直接放入有效的 dlssnr-amd/dlssnr.bin。本包不附带上述任一文件。

已知问题

• lmxxf 在部分游戏场景下的闪退问题（已知包括《鸣潮》、《异环》、《伊莫》等）正在与上游协同排查修复，当前遇到闪退推荐尝试 Mochizuki 后端。
• 可能无法与 ReShade 插件兼容。

特别感谢：各位 Bilibili 用户的测试与反馈意见。

---

1.10.1 (compared to 1.10.0)

• Fixed an issue where the Mochizuki backend failed to run (shader compilation error) on AMD driver 26.9.2 and newer. Special thanks to @MatheusFerreiraS for providing the fix.

Daniel backend

Download the runtime, weights, and nvngx_dlssnr.dll 310.8.0 separately from the Daniel releases page. Daniel 0.6.0 remains supported.

lmxxf backend

Weight files: lmxxf upstream repository or download mirror.

Mochizuki backend

Model extraction requires your own supported nvngx_dlssnr.dll 310.8.0 (same as the Daniel backend) and Python 3.10+, or an existing valid dlssnr-amd/dlssnr.bin. This package contains neither file.

Known Issues

• lmxxf crashes in certain game scenarios (known cases: Wuthering Waves, Neverness to Everness, Yimo, etc.) are being investigated in collaboration with upstream. Switching to the Mochizuki backend is recommended if crashes occur.
• May conflict with ReShade.

Special Thanks: Thank you to all Bilibili users for your testing and feedback.

This archive does not include NVIDIA proprietary binaries, the danielblnc installer, or unauthorized model weights.

See the included README for configuration details. Bug reports with logs are welcome.

Full Changelog: v1.10.0...v1.10.1