github Likely7/Veyra-NRVideo v2.0.6
Veyra 2.0.6 · VFG performance, AMD NR & Blackmagic capture

latest releases: v2.2.0, v2.2.0-sources, v2.1.9...
one month ago

Veyra 2.0.6

本次重点:VFG 补帧吞吐优化、AMD NR 开启后结果被后续 FSR 覆盖的修复,以及 Blackmagic/DeckLink 采集兼容。提供 NVIDIA / AMD 两个完整便携包;应用版本为正式版,NR 与社区/解锁运行组件继续保留各自实验边界。

VFG:修复“有余量却上不去”的一类瓶颈

原先 VFG 的 CUDA/SDK 提交工作会占住呈现线程,GPU 性能球有余量也可能错过呈现时机。现在提交工作进入有上限的独立 worker,每个输出槽显式区分等待、成功与失败;只有有效完成的结果参与呈现,失败仍保留源帧。GPU 同步、停机回收和离线导出完成判定保持完整。

只调整 NVIDIA VFG,不改 DLSS、XeSS 或 FSR 补帧算法,也不通过提高默认 GPU 优先级获得这次数据。队列最多 14 个任务,覆盖两组 8X 工作;这不是无界缓存。运行库缺失和生成失败仍有回退,不把未完成纹理当成功帧。

匹配性能对照

RTX5070 12GB / 驱动616.56;GTA VI 4K30 视频输入,原版单层 NR、内部1080p、强度1、风格0、零运动,自动调控/抗闪烁/SR/HDR/调色关闭;VFG仍处理4K。窗口1280×800、自动显示同步、输出不限帧,实际普通GPU优先级。每轮预热5秒、统计20秒,无并发GPU任务。下表是同次优化验收的相邻匹配对照中位数,计量为软件呈现/提交 FPS。

VFG 档位 优化前 优化后 提升 CPU 提交滚动 P95 的中位数
中等 4X 80 FPS 120 FPS +50.0% 23.450 → 1.198 ms
高 2X 35.5 FPS 60 FPS +69.0% 30.940 → 1.079 ms

同一候选程序把中等4X强制切回同步提交后为76.5 FPS,正常异步提交为120 FPS,支持本次瓶颈定位。更完整矩阵中:低档2–8X达到对应60–240 FPS目标,中等2–5X达到60–150 FPS,高档仅2X达到60 FPS。更高倍率仍受计算/带宽等预算限制;不能推断“高档8X也能跑满”。CPU提交减少不等于推理耗时、屏幕延迟或显存占用同比减少。

既有优化验收覆盖3组GPU测试各913检查、168项同步/异步输出hash一致、设置331检查、21种720p导出及4K NR/VFG导出,另有切换/暂停/seek/全屏/取消/失败回退检查。2.0.6另进行新构建的针对性回归,结果见发布验收。详细配置和原始数据见VFG报告。

AMD NR:修复开启执行但画面没有变化

在“NR先处理 → FSR超分”的链路中,NR已有正常执行记录,后续FSR却可能再次读入原图,覆盖NR结果;普通缩放的首帧/reset路径也存在同类输入选择问题。现在统一读取当前已完成的处理阶段输出,覆盖文件播放、串流共用图及相关导出路径。此修复不修改NR模型、强度参数、HIP运行库或FSR算法。

此前定向复现6个用例失败且最大像素码值偏差156;修复后20个图用例全部通过,最大偏差0、D3D12验证错误0,包含NR先行、FSR、普通缩放、首帧/reset、多层、运动/时间历史及4K导出图。该证据来自RTX5070上真实FSR3.1.5与仅用于测试的确定性NR provider;AMD HIP、FSR4.1 ML与用户Xbox实机仍需复测,不以替代测试冒充实卡效果验收。详见AMD修复报告。

Blackmagic / DeckLink 兼容

  • 补齐Blackmagic WDM可能需要的设备medium/上游crossbar连接,不替用户改变输入路由。
  • 支持HDYC按UYVY解包,并在无明确元数据/手工覆盖时采用BT.709;保留显式颜色设置优先级。
  • 按设备返回的精确格式协商,保留VideoInfo2扫描/交错字段,减少格式看似相同但无法连接的问题。
  • 采集断开时可打开设备/输入驱动属性,选择正确HDMI等输入;增加有限频率的黑场采样诊断,帮助区分已收到黑画面与完全没帧。

来源、固定提交与LGPL声明已记录在THIRD_PARTY_NOTICES。现有回归包括213项采集/COM/颜色检查、124项GPU颜色检查和本机USB卡约60FPS回调/重连检查。本机没有Blackmagic卡,尚不能确认反馈者真实HDMI信号已恢复;列表中的两个名称也不一定是两张卡。 详见采集报告。

下载、更新与反馈

  1. 按显卡选择NVIDIA或AMD完整便携包,在新目录完整解压;保留2.0.5方便回退,不把两个厂商包覆盖混装。
  2. 本次不升级运行组件/模型;patched FFmpeg、许可证和对应依赖源码继续随包提供。校验下载请使用SHA256SUMS.txt。应用及依赖源码见Release独立资产;重编译/静态依赖重链接见构建说明。
  3. 4群已满,README和本Release更新为用户提供的交流群5二维码,图片标注2026-10-14前有效;微信赞助二维码、Discord与Ko-fi继续保留。
  4. 反馈请附显卡/驱动、输入来源、NR/补帧运行版本、处理顺序与日志。公开日志前请隐藏账号/会话信息。

已知:NVIDIA持续显存增长尚未定位/修复;部分高倍率VFG仍达不到目标帧率;AMD实卡HIP/FSR4.1、Blackmagic真实输入、所有设备的长期稳定性未全部验收。XeSS补帧/节点离线导出仍不支持。Xbox日志中的独立硬解回退问题未在本次修复范围内。


English release notes

2.0.6 improves NVIDIA VFG scheduling, fixes the NR-first output handoff to FSR/scaling, and adds Blackmagic/DeckLink capture compatibility. Separate complete NVIDIA and AMD portable packages are available. The application release is stable; experimental NR/community/unlock components retain their existing status.

VFG: bounded worker submission removes CUDA/SDK CPU work from the presentation thread. Output slots explicitly track pending/success/failure, with source-frame fallback and complete teardown synchronization. Other frame-generation algorithms and runtime binaries are unchanged. Matched adjacent RTX5070 tests with GTA 4K30, one original NR layer at internal1080p and normal GPU priority improved Medium4X from80 to120 software-present FPS (+50.0%) and High2X from35.5 to60 (+69.0%). Median rolling CPU-submission P95 changed23.450→1.198ms and30.940→1.079ms respectively. Each run used5s warmup+20s observation, VFG at4K, no competing GPU workload. These are conditional submission measurements, not screen refresh, end-to-end latency or VRAM savings. Higher multipliers remain budget-limited. Full settings/evidence and fresh release checks are linked above.

NR → FSR/scaling: the later stage could read the original image despite NR executing successfully, making the NR toggle appear ineffective. Both FSR and relevant scaling/reset paths now consume the completed stage output. Six reproduced failures became20 passing graph cases with zero marker error/D3D12 validation errors. This uses real FSR3.1.5 and a deterministic test NR provider on RTX5070; AMD HIP/FSR4.1 ML and Xbox hardware confirmation remain pending. The test provider is never shipped.

Capture: Blackmagic WDM upstream connections, HDYC/UYVY with a BT.709 fallback, exact driver format negotiation including scan/interlace metadata, disconnected driver-property dialogs, and bounded black-frame sampling diagnostics. No local Blackmagic hardware was available; this is compatibility implementation plus regression evidence, not confirmed physical HDMI acceptance.

Extract the appropriate vendor package into a new folder. Runtime/model/codec identities and patched FFmpeg are preserved from2.0.5; application and dependency source archives, notices and checksums accompany the release. Group4 is full; the supplied Group5 QR replaces the current community QR and is valid before2026-10-14. Existing donation QR, Discord and Ko-fi remain. NVIDIA sustained VRAM growth is unresolved; higher VFG multipliers and long-running cross-device stability are not universally validated. XeSS FG/node offline export remains unsupported.


支持与反馈 / Support & feedback

Join Veyra on Discord    Patreon 主页 / Creator page   

Patreon 主页查看作者动态;☕ 咖啡按钮直达 US$5 赞助页面。
Visit the creator page for updates, or open the US$5 coffee support page.

微信赞助      Veyra 交流群 5

左:微信赞助(自愿,不影响功能);右:交流群。群码按图片标注于 2026-10-14 前有效,过期请查看仓库更新。

Left: optional donation; right: community group. QR valid before 2026-10-14.

Don't miss a new Veyra-NRVideo release

NewReleases is sending notifications on new releases.