English
- Apply the accepted Uhra TAA policy to the normal TAA path: 0.5 jitter scale, stationary motion snapping, stationary color clipping and multi-surface history, with RGBA8 history at
31/33. Experimental FP16 history and moving bilinear fallback remain off. - Enable experimental geometric motion-vector replay for TAA by default, while retaining
LO_MV_ENABLE=0as a comparison switch. Repeated instances are paired by stable submission order so adjacent-frame Nth-to-Nth motion matching remains intact. - Reuse exact-content index fingerprints in the index cache while retaining complete source-byte verification. In the same Uhra Vulkan 4K internal/output scene on an RTX 5080, hidden muted A-B-A-B captures without pacing measured 60.34/59.00 FPS for the candidate, versus 54.61 FPS for a separate Release build and 54.57 FPS for the former RelWithDebInfo main binary.
- Validation includes MV audit steady tracked/matched/replay 988 with failed 0,
LoMotionVectorTest71 checks andLoVertexCacheTest3,668,948 checks. User foreground review of the same Uhra steel-frame scene found image quality acceptable at about 60 FPS. - Scope remains bounded to the tested Uhra scene, Vulkan and the local RTX 5080. The 1080p-internal to 4K moving-camera limitation, broader scene coverage and D3D12 replay PSO creation follow-up remain open.
简体中文
- 将已接受的 Uhra TAA 策略应用到正常 TAA 路径：0.5 抖动幅度、静止运动 snap、静止颜色裁剪和多表面 history，使用 RGBA8 history 与
31/33权重。实验性 FP16 history 和 moving bilinear fallback 仍关闭。 - TAA 默认启用实验性几何运动矢量 replay，保留
LO_MV_ENABLE=0对照开关。重复实例继续按稳定提交顺序配对，保持相邻帧 Nth-to-Nth 运动匹配。 - index cache 复用 exact-content index fingerprint，同时保留完整源字节验证。在 RTX 5080 的 Vulkan、Uhra 4K 内部／输出同一场景中，隐藏静音、无 pacing 的 A-B-A-B 对照测得候选 60.34/59.00 FPS；独立 Release 构建为 54.61 FPS，之前的 RelWithDebInfo 主程序为 54.57 FPS。
- 验证包括 MV audit steady tracked/matched/replay 为 988、failed 为 0，
LoMotionVectorTest71 项和LoVertexCacheTest3,668,948 项。用户在同一 Uhra 钢架场景前台观察，确认画质可接受、约 60 FPS。 - 范围限定为已测试的 Uhra 场景、Vulkan 和本机 RTX 5080。1080p internal 到 4K output 的移动相机限制、更广场景覆盖和 D3D12 replay PSO 创建后续工作仍开放。