MagicBeanDev/Longshot

Watch

Fork 98

Game: PsychoV35 warm/cool flicker with camera movement and area transitions #344

New issue

Open

opened 2026-11-14 02:50:52 +00:00 by futondestroyer · 3 comments

futondestroyer commented

2026-11-14 02:50:52 +00:00

Member

Longshot version

nightly-77-9f64464b (source 6f23b63d), from the supplied player log.

Channel

Nightly

What kind of problem

Graphics: camera-motion colour flicker with PsychoV35.

What happened

A player reports repeatable flickering while moving the camera in the Lon Lon Ranch alley. It does not occur with the Vanilla tone mapper and still occurs after a fresh reboot. They had tested for less than an hour and had not noticed it elsewhere.

The supplied 8.17-second HDR video shows abrupt warm/cool shifts across the ground between nearly identical camera views. This is not the earlier missing-opt-in fallback: the supplied log confirms the PsychoV35 RGBA16F route, an HDR DLAA feature and sustained NGX frame generation.

Where in the game

Lon Lon Ranch alley, child Link, stationary position with camera movement. The reporter can reproduce it at will there.

How often

Repeatable in that location according to the reporter; reach elsewhere is not established.

Graphics card and driver version

NVIDIA GeForce RTX 5090, driver 621.24.

CPU, RAM and operating system

Windows; about 32 GB RAM reported in the log. Exact CPU model and OS build not supplied.

Changes from the defaults

3840x2160 output, Native HDR, PsychoV35, paper white 200 nits / peak 1000 nits, DLAA preset K, NVIDIA frame generation 2x: the active feature logs one made frame per set. The advertised maximum of five made frames is device capability, not the selected multiplier. Other overrides/overlays are not fully established from the available evidence.

longshot_log.txt

The full log and video were supplied privately for this investigation and have not been uploaded. Relevant non-personal excerpts:

longshot nightly-77-9f64464b (source 6f23b63d)
HDR: an HDR10 surface (format 64), paper white 200 nits, peak 1000 nits
NVIDIA DLSS: feature 1 made ... 3840x2160 ... mode 5, preset K, HDR colour, linear (the game's native HDR)
PsychoV35 DLSS: RGBA16F linear route; scene format 122, network input 97, output 97
DLSS Frame Gen (NGX): of the last 600 presents 600 made by DLSS-G ... 0 without new inputs ... 0 asked more than 5 frames

Investigation status

Claimed by this author. Severity S2 (visual corruption); reach includes repeated ranch camera flashes on RTX4070 and the original RTX5090 report, plus operator-reported flashes elsewhere and around Lake Hylia/Field transfers. Confidence1.0 in the reproduced visual defect; root cause remains unconfirmed.

The operator confirmed the captured preview, then manually reproduced repeated flashes on the original514591a3 renderer with FG Off; the complete recording was preserved. The RGBA16F route stayed active at the ranch flashes. The separate hue-continuity candidate258040c2 also shows the symptom and is not a solution for this report. No fix PR is open. Newer nightly-127-cf605c85 gameplay also reports the symptom; its occasional format fallbacks are a separate lead. See #344 (comment) for settings, recorded evidence, limitations and the shader-permutation diagnostic in progress.

Related opt-in route: #108. This is separate from the DLSS sizing performance change #311. No game files, saves, compiled artifacts or raw dumps are attached.

### Longshot version nightly-77-9f64464b (source 6f23b63d), from the supplied player log. ### Channel Nightly ### What kind of problem Graphics: camera-motion colour flicker with PsychoV35. ### What happened A player reports repeatable flickering while moving the camera in the Lon Lon Ranch alley. It does not occur with the Vanilla tone mapper and still occurs after a fresh reboot. They had tested for less than an hour and had not noticed it elsewhere. The supplied 8.17-second HDR video shows abrupt warm/cool shifts across the ground between nearly identical camera views. This is not the earlier missing-opt-in fallback: the supplied log confirms the PsychoV35 RGBA16F route, an HDR DLAA feature and sustained NGX frame generation. ### Where in the game Lon Lon Ranch alley, child Link, stationary position with camera movement. The reporter can reproduce it at will there. ### How often Repeatable in that location according to the reporter; reach elsewhere is not established. ### Graphics card and driver version NVIDIA GeForce RTX 5090, driver 621.24. ### CPU, RAM and operating system Windows; about 32 GB RAM reported in the log. Exact CPU model and OS build not supplied. ### Changes from the defaults 3840x2160 output, Native HDR, PsychoV35, paper white 200 nits / peak 1000 nits, DLAA preset K, NVIDIA frame generation 2x: the active feature logs one made frame per set. The advertised maximum of five made frames is device capability, not the selected multiplier. Other overrides/overlays are not fully established from the available evidence. ### longshot_log.txt The full log and video were supplied privately for this investigation and have not been uploaded. Relevant non-personal excerpts: ```text longshot nightly-77-9f64464b (source 6f23b63d) HDR: an HDR10 surface (format 64), paper white 200 nits, peak 1000 nits NVIDIA DLSS: feature 1 made ... 3840x2160 ... mode 5, preset K, HDR colour, linear (the game's native HDR) PsychoV35 DLSS: RGBA16F linear route; scene format 122, network input 97, output 97 DLSS Frame Gen (NGX): of the last 600 presents 600 made by DLSS-G ... 0 without new inputs ... 0 asked more than 5 frames ``` ### Investigation status Claimed by this author. Severity S2 (visual corruption); reach includes repeated ranch camera flashes on RTX4070 and the original RTX5090 report, plus operator-reported flashes elsewhere and around Lake Hylia/Field transfers. Confidence1.0 in the reproduced visual defect; root cause remains unconfirmed. The operator confirmed the captured preview, then manually reproduced repeated flashes on the original514591a3 renderer with FG Off; the complete recording was preserved. The RGBA16F route stayed active at the ranch flashes. The separate hue-continuity candidate258040c2 also shows the symptom and is not a solution for this report. No fix PR is open. Newer nightly-127-cf605c85 gameplay also reports the symptom; its occasional format fallbacks are a separate lead. See https://git.magicbean.dev/MagicBeanDev/Longshot/issues/344#issuecomment-17670 for settings, recorded evidence, limitations and the shader-permutation diagnostic in progress. Related opt-in route: https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/108. This is separate from the DLSS sizing performance change https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/311. No game files, saves, compiled artifacts or raw dumps are attached.

futondestroyer commented

2026-11-14 03:28:50 +00:00

Author

Member

Local investigation update (2026-11-14 UTC):

The local operator also saw flicker while manually positioning Link, with a tentative correlation to looking toward the sky. That manual interval was not recorded. It must not be treated as a negative reproduction just because subsequent automatic captures did not establish the same symptom.
The manual session was configured with frame generation Off: the first captured option transition is frame_generation 0 -> 0; the first NGX FG feature is made only after the later automated 0 -> 2 transition. This weakens an FG-only explanation, although a recorded repeat is still needed. The reporter's independent FG-off comparison remains pending.
Completed matching saved-daylight-alley original/candidate runs, each with PsychoV35/Vanilla and FG Off/2x: RTX 4070, NVIDIA 621.01, Windows, 2560x1440, Native HDR 200/1000 nits, DLAA K. Original: 514591a3. Candidate: identical host/base with only psychov35.h changed, renderer SHA-256 044fd7d6abada480ca29f6f2de673137bc89499e3ab522eded3c1f7da240a52d. This is not the reporter's 4K/5090 configuration. Short camera sweeps do not yet establish the video's warm/cool flash or its removal; motion-edge detections also appear on Vanilla and are not a useful pass/fail result alone. No performance conclusion is drawn from these capped runs.
The isolated numerical defect is reproducible: last native-fit channel crosses saturation near 7.241657. For scene RGB (10,9,7.2415) -> (10,9,7.2418), gain .03 and peak 5, the original output changes from (1.23764539,1.17456961,.880141616) to (1.28815556,1.15565848,.918046117). Draft attenuation gives (1.28815031,1.15565991,.918001413) -> (1.28815556,1.15565848,.918046117). This is a mathematical continuity fix, not yet a demonstrated fix for this report.
Candidate numerical test: 3153 colours and 72 saturation-boundary pairs pass; original code fails the new boundary assertion. Renderer-only pack builds and links with zero compiler failures, and its scene-tone-map replacement runs in-game. Full game-free suite: 516 pass / 4 fail / 34 skip. All four failures also reproduce on an untouched 76699aad control: test_oot_state_tap, test_pad_layout, test_release, test_tested_configurations.

Next reproduction should explicitly capture the sky-to-ground transition and exposure settling, including the manual portion. Our game instances are closed and the shared lock released for the other chat's 23:30 Halifax build slot. No PR is claimed ready, and this issue remains open.

Source-only numerical candidate: https://git.magicbean.dev/futondestroyer/Longshot/src/branch/codex/psychov35-camera-flicker (258040c2). No PR yet; the camera-flicker root cause remains unconfirmed.

Local investigation update (2026-11-14 UTC): - The local operator also saw flicker while manually positioning Link, with a tentative correlation to looking toward the sky. That manual interval was not recorded. It must not be treated as a negative reproduction just because subsequent automatic captures did not establish the same symptom. - The manual session was configured with frame generation Off: the first captured option transition is `frame_generation 0 -> 0`; the first NGX FG feature is made only after the later automated `0 -> 2` transition. This weakens an FG-only explanation, although a recorded repeat is still needed. The reporter's independent FG-off comparison remains pending. - Completed matching saved-daylight-alley original/candidate runs, each with PsychoV35/Vanilla and FG Off/2x: RTX 4070, NVIDIA 621.01, Windows, 2560x1440, Native HDR 200/1000 nits, DLAA K. Original: 514591a3. Candidate: identical host/base with only `psychov35.h` changed, renderer SHA-256 `044fd7d6abada480ca29f6f2de673137bc89499e3ab522eded3c1f7da240a52d`. This is not the reporter's 4K/5090 configuration. Short camera sweeps do not yet establish the video's warm/cool flash or its removal; motion-edge detections also appear on Vanilla and are not a useful pass/fail result alone. No performance conclusion is drawn from these capped runs. - The isolated numerical defect is reproducible: last native-fit channel crosses saturation near 7.241657. For scene RGB (10,9,7.2415) -> (10,9,7.2418), gain .03 and peak 5, the original output changes from (1.23764539,1.17456961,.880141616) to (1.28815556,1.15565848,.918046117). Draft attenuation gives (1.28815031,1.15565991,.918001413) -> (1.28815556,1.15565848,.918046117). This is a mathematical continuity fix, not yet a demonstrated fix for this report. - Candidate numerical test: 3153 colours and 72 saturation-boundary pairs pass; original code fails the new boundary assertion. Renderer-only pack builds and links with zero compiler failures, and its scene-tone-map replacement runs in-game. Full game-free suite: 516 pass / 4 fail / 34 skip. All four failures also reproduce on an untouched 76699aad control: test_oot_state_tap, test_pad_layout, test_release, test_tested_configurations. Next reproduction should explicitly capture the sky-to-ground transition and exposure settling, including the manual portion. Our game instances are closed and the shared lock released for the other chat's 23:30 Halifax build slot. No PR is claimed ready, and this issue remains open. Source-only numerical candidate: https://git.magicbean.dev/futondestroyer/Longshot/src/branch/codex/psychov35-camera-flicker (258040c2). No PR yet; the camera-flicker root cause remains unconfirmed.

futondestroyer commented

2026-11-14 08:01:17 +00:00

Author

Member

Local reproduction is now recorded and confirmed by the operator. The test used the original 514591a3 renderer, PsychoV35, Native HDR (200/1000 nits), DLAA K at2560x1440 on an RTX4070/NVIDIA621.01, with frame generation Off throughout. The operator deliberately repeated the ranch camera movement and then moved toward Lake Hylia; the recording started before manual input and has12874 complete records with no partial tail.

Several ranch transitions are clear in adjacent frames: for example frame2724 has a median inter-frame RGB change of approximately(-8.29,-6.15,-2.09) across the diagnostic blocks, with75 of98 blocks changing red-minus-blue by more than2 code values. This is a broad warm/cool colour change, not a motion-edge-only detection. Pause-menu transitions are excluded from that example. This corrects the earlier lack of a recorded reproduction.

The separate hue-continuity candidate258040c2 also produced a captured flash with FG Off; the operator confirmed its preview matches the symptom, although they saw repeated events. It is NOT a demonstrated fix for this issue, and no PR is open for it.

The original manual run stayed on the RGBA16F route throughout gameplay: there is no packed/linear route switch at those ranch flashes. A newer supplied gameplay log from nightly-127-cf605c85 (RTX4070,3440x1440,Native HDR,DLAA K and2x FG) independently has several unexpected-format fallbacks and packed/linear route changes, including a late one at frame173796. The operator changed settings during early ranch testing but not around the later Lake Hylia/Field transfer. The log has no event timestamp tying that late fallback to a specific visible flash, so it is a separate lead rather than an established explanation of all events.

A concrete shader-matching lead is now being checked: program7370b01162e62427 is logged as left on the game's curve because the fit input is not times a constant, whereas672e593076ac390a receives PsychoV35. A renderer-only diagnostic pack will correlate tone-map permutation use with captured frames. This hypothesis is not yet established. No game artifacts or media are uploaded; recordings and full logs remain local.

Local reproduction is now recorded and confirmed by the operator. The test used the original 514591a3 renderer, PsychoV35, Native HDR (200/1000 nits), DLAA K at2560x1440 on an RTX4070/NVIDIA621.01, with frame generation Off throughout. The operator deliberately repeated the ranch camera movement and then moved toward Lake Hylia; the recording started before manual input and has12874 complete records with no partial tail. Several ranch transitions are clear in adjacent frames: for example frame2724 has a median inter-frame RGB change of approximately(-8.29,-6.15,-2.09) across the diagnostic blocks, with75 of98 blocks changing red-minus-blue by more than2 code values. This is a broad warm/cool colour change, not a motion-edge-only detection. Pause-menu transitions are excluded from that example. This corrects the earlier lack of a recorded reproduction. The separate hue-continuity candidate258040c2 also produced a captured flash with FG Off; the operator confirmed its preview matches the symptom, although they saw repeated events. It is NOT a demonstrated fix for this issue, and no PR is open for it. The original manual run stayed on the RGBA16F route throughout gameplay: there is no packed/linear route switch at those ranch flashes. A newer supplied gameplay log from nightly-127-cf605c85 (RTX4070,3440x1440,Native HDR,DLAA K and2x FG) independently has several unexpected-format fallbacks and packed/linear route changes, including a late one at frame173796. The operator changed settings during early ranch testing but not around the later Lake Hylia/Field transfer. The log has no event timestamp tying that late fallback to a specific visible flash, so it is a separate lead rather than an established explanation of all events. A concrete shader-matching lead is now being checked: program7370b01162e62427 is logged as left on the game's curve because the fit input is not times a constant, whereas672e593076ac390a receives PsychoV35. A renderer-only diagnostic pack will correlate tone-map permutation use with captured frames. This hypothesis is not yet established. No game artifacts or media are uploaded; recordings and full logs remain local.

futondestroyer changed title from ~~Game: PsychoV35 warm/cool flicker with camera movement in Lon Lon Ranch alley~~ to Game: PsychoV35 warm/cool flicker with camera movement and area transitions

2026-11-14 08:01:47 +00:00

futondestroyer referenced this issue

2026-11-14 09:14:25 +00:00

fix(vk): apply PsychoV35 to the adapted scene permutation #400

futondestroyer commented

2026-11-14 09:19:10 +00:00

Author

Member

Confirmed cause of the recorded ranch palette flashes: the two scene tone-map permutations switched between PsychoV35 and an unmodified vanilla curve. The adapted permutation includes a common exp2 exposure multiplier that the matcher did not recognize. In a recorded diagnostic control, all 45 stable-play switches matched the expected warm/cool direction at a fixed capture offset. PR #400 now recognizes that exposure shape while preserving the complete gain and existing guards. Two interleaved original/fixed pairs gave 45 and 37 large chroma-step detections before, zero in both fixed runs; both scene permutations now log PsychoV35. Full checks have the same six failures as clean main (521 pass/6 fail/34 skip). The PR is WIP for Linux compile coverage T-S724 and permission to attach the prepared comparison. No merge has been performed. The separate transition-format fallback from the newer gameplay log is not claimed fixed; please retest the broader transfer symptom once this is available.

Confirmed cause of the recorded ranch palette flashes: the two scene tone-map permutations switched between PsychoV35 and an unmodified vanilla curve. The adapted permutation includes a common exp2 exposure multiplier that the matcher did not recognize. In a recorded diagnostic control, all 45 stable-play switches matched the expected warm/cool direction at a fixed capture offset. PR https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/400 now recognizes that exposure shape while preserving the complete gain and existing guards. Two interleaved original/fixed pairs gave 45 and 37 large chroma-step detections before, zero in both fixed runs; both scene permutations now log PsychoV35. Full checks have the same six failures as clean main (521 pass/6 fail/34 skip). The PR is WIP for Linux compile coverage T-S724 and permission to attach the prepared comparison. No merge has been performed. The separate transition-format fallback from the newer gameplay log is not claimed fixed; please retest the broader transfer symptom once this is available.