[Arkex-ai]

Summary

Invalid camera reprojection disables motion vectors but can leave temporal history without a valid mapping. Reset that history in the same frame.

Technical changes

• Reject non-finite transforms and float overflow before conversion.
• Reset shared DLSS/FSR 4 and separate FSR 3.1 consumers.
• Preserve first-frame, gap and size rules, API13 and main's audio, filtering, LOD bias and capture defaults.

Validation

Actual reprojection, reset policies and consumers pass singular/non-finite, overflow, recovery, gap and ultrawide projection cases.

Head 19b81b7, main 8c92fe3: full local suite 301 passed / 3 failed / 21 skipped, versus 299 / 4 / 21 on the exact base. Three shared failures: Windows symlink privilege, missing zstd CLI in the new bake fixture, and the local Windows-header harness. Baseline also timed out in frame-wait; its identical production block and fixtures passed all18 cases in focused reruns on both base and candidate. Original counts are preserved.

Matching NGX host/backend build passed with officially regenerated runtime. Windows builder, Linux AppImage and NGX CIs passed. Source CI failed: Windows cannot load libzstd and hits a logging timing assertion; Linux rejects the bake compiler's C++23 probe. Exact upstream baseline CI has those failures too. This run additionally timed out in Windows bake_skeleton (600s). No temporal check failed; the source workflow remains red. Bounded native smoke on RTX5070Ti passed: copied Kokiri checkpoint, official DLSS310.9.16/preset K loaded and evaluated, window/fullscreen, pause/resume and exit0; eight3440x1440 engine captures. Internal upscale dimensions stayed fixed. SDK recreation, render-scale transitions, quality and performance comparisons remain untested.

[hckrman101]

Looks good. The reprojection now rejects non-finite inputs and float-overflowing outputs, and both the DLSS and FSR paths promote a failed reprojection to a history reset, which is the right thing when motion vectors are off. Existing first-frame, frame-gap and FSR size-change resets are untouched, and the new test is registered in run_all.py. CI is green.

Nit: the first-frame/gap case already forces reset, so the new line only matters when the matrix math fails, which is exactly the intent. Nothing to change.

[hckrman101]

Heads-up: this conflicts with main now (a lot landed today). Could you merge main into the branch (or rebase) and resolve the conflicts? Thanks!

[Arkex-ai]

Refreshed on main 8c92fe3 at 19b81b7. GitHub reports the branch as mergeable. Matching NGX host/backend build and bounded native checkpoint/window/fullscreen/pause-resume smoke passed, with exit0 and original saves preserved. SDK recreation and render-scale transitions are still untested. Three fork CIs passed; source remains red on bake/logging checks also failing in the exact upstream baseline, plus a Windows bake timeout in this run. No temporal check failed. The PR description links both CI reports.

[hckrman101]

Re-reviewed after the rebase onto main. The change is the same as the one I approved, now applied to the shared chain in vk_temporal.cpp (DLSS/FSR 4) and to vk_fsr3_inputs.cpp (FSR 3.1): a failed reprojection forces reset, and reprojection() rejects non-finite inputs and float-overflowing results. The conflict I flagged is resolved and GitHub reports the branch mergeable.

I checked that the reset reaches each consumer: f.reset = reset in vk_temporal.cpp feeds InReset in vk_dlss.cpp and d.reset in vk_fsr4.cpp, and FSR 3.1 sets f.reset = reset directly. The new test splices those production fragments into its fixture, so it follows the code.

CI: the Windows builder, DLSS/FSR 4 link check, AppImage, actionlint and the SPIR-V/F2I jobs pass. Python source checks fail on both OSes, but main's own source-regressions run at 8c92fe3 is red too, and the author documents the same bake and logging failures there. I found nothing temporal in them.

Nit: SDK recreation and render-scale transitions are untested, as the author says. Nothing to change here.