Home

Cutscenes that crash, or stay black, in Windows games running under CrossOver on Apple Silicon — what causes it, which games are affected, and how to find out about a game that is not listed yet.

This is for CrossOver Preview

Specifically crossover-preview-arm64-20260821. That is where every title here is measured, and it is the only configuration this project supports. Three of the nine also run on stable 26.3; treat that as a bonus rather than a promise.

The tooling lives in MacGameVideoFix.

The two failure modes

They look different and have nothing in common except the symptom.

The crash. Unreal's Electra media player asks every D3D12 resource for ID3DDestructionNotifier and uses the answer without checking whether it got one. Apple's D3DMetal does not implement that interface, so the first VP9 frame dereferences a null vtable and the game dies. H.264 and H.265 can avoid the buffer pool through a CVar; VPx has no equivalent, so VP9 on D3D12 has no way out through configuration.

→ Fixed. See Mortal Shell 2.

The black screen. The game reaches the cutscene, shows nothing, and does not crash. Nothing returns an error, so there is no crash log and nothing to grep for — the failure has to be traced through the code.

→ Fixed on DYNASTY WARRIORS: ORIGINS, where it turned out to be five separate faults in a row, ending in a frame that decoded correctly and had no way to reach the renderer that draws it.

A third mode has nothing to do with video at all. Both Life is Strange titles run fine and then freeze, anywhere, because Unreal walks the GPU's memory nodes and D3DMetal never tells it to stop. It is listed here because it is the same toolkit, not because it is the same problem.

Is my game affected?

The thing that matters is what the cutscenes are encoded as, what box they are in, and which API plays them. survey-games.sh reports all three for a game folder:

diagnostics/survey-games.sh "/path/to/steamapps/common/<Game>"

Reading the output:

• VP9 + Unreal — likely the crash, but confirm it before patching. No static scanner for Electra's 12000 version check ships here: the code that knows the pattern is the runtime patch itself, which reports how many sites it found once it is installed. A count of zero means this bug is not present in that build, whatever else may be wrong. See Diagnosing a new game.
• VP9 + anything else — possible, but a different mechanism each time.
• Bink (.bik / .bk2) — Bink ships its own decoder and never touches Media Foundation or D3D video. Not affected by any of this.
• H.264 — decoded by CrossOver on its own, on stable and on Preview, with nothing patched. A working decoder is not the whole story: Beast of Reincarnation decodes H.264 and still showed nothing, because CrossOver withholds NV12 from the format list on macOS and Electra accepts nothing else. If the sound plays and the picture does not, start at Beast of Reincarnation.

The container is worth as much as the codec. The codec says whether anything can decode the file; the container says whether anything can open it, and on stable CrossOver that is where WebM stops — see Games.

Two caveats on the survey. It reads Unreal .pak indexes but only version 11 unencrypted ones, so a title using anything else reports zero videos when it may have hundreds. And a game that packs its movies in a proprietary archive is invisible to it — 0 means "none found loose or in a readable pak", never "no videos".

Games

Game	Engine	Symptom	Fix	Backend	DX	GPTK	CrossOver	Status
Mortal Shell 2	Unreal Engine 5.6.1	Crash on the first cutscene	Runtime patch, 4 sites	D3DMetal	12	3.0 and 4.0b2	26.3 and Preview	Fixed
Life is Strange: Reunion	Unreal Engine 5	Freezes after a while, anywhere	DXGI node guard	D3DMetal	12	4.0b2 only -- 3.0 crashes it	26.3 and Preview	Fixed
Life is Strange: Double Exposure	Unreal Engine 5	Freezes after a while, anywhere	DXGI node guard, same DLL	D3DMetal	12	4.0b2 only -- 3.0 crashes it	26.3 and Preview	Fixed
DYNASTY WARRIORS: ORIGINS	Koei Tecmo, in-house	Cutscene runs with sound, picture black	Video bridge, D3D11 to D3D12	D3DMetal	12	4.0b2	26.3 and Preview	Fixed
Beast of Reincarnation	Unreal Engine 5	Startup video plays with sound, no picture	NV12 restored, Electra forced to software	D3DMetal	12	3.0 and 4.0b2	26.3 and Preview	Fixed
Persona 5 Strikers	Koei Tecmo, in-house	Video never starts; sound only	Staged VC-1 codec, and a D3D9 to D3D11 bridge	DXMT	11	3.0 and 4.0b2	26.3 and Preview	Fixed
Nioh	Koei Tecmo, in-house	Cutscene refuses to play, then crashes	Staged WMV3 codec, and the same D3D9 to D3D11 bridge	DXMT	11	4.0b2	26.3 and Preview	Fixed
Nioh 2	Koei Tecmo, in-house	Cutscene refuses to play, then crashes	Same codec and same bridge as Nioh, unchanged	DXMT	11	4.0b2	26.3 and Preview	Fixed
Nioh 3	Koei Tecmo, in-house	Failed to play movie	The DYNASTY WARRIORS bridge, unchanged	D3DMetal	12	4.0b2	26.3 and Preview	Fixed
Wo Long: Fallen Dynasty	Koei Tecmo, in-house	Cutscene runs with sound, picture black	The DYNASTY WARRIORS bridge, unchanged	D3DMetal	12	4.0b2	26.3 and Preview	Fixed
NieR Replicant ver.1.22474487139	Toylogic, in-house	Crashes when the first video starts	Software decode, and the frame written into the game's target	D3DMetal	11	4.0b2	26.3 and Preview	Fixed
KINGDOM HEARTS Dream Drop Distance	Square Enix, in-house	Cutscene runs with sound, picture solid green	Software decode, and the luma and chroma planes written into the game's own textures	D3DMetal	11 + 12	4.0b2	Preview -- not tried on 26.3	Fixed
KINGDOM HEARTS HD 1.5+2.5 ReMIX	Square Enix, in-house	Cutscene runs with sound, picture solid green	The Dream Drop Distance fix, unchanged -- six executables, same route	D3DMetal	11 + 12	4.0b2	Preview -- not tried on 26.3	Fixed
TMNT: Splintered Fate	Rebirth, in-house	Opens a window, then closes silently	A guard on the D3D12 call that ends the process instead of failing	D3DMetal	12	4.0b2	26.3 and Preview	Fixed
Tormented Souls 2	Unreal Engine 5	Fatal error before the first frame	16:9 modes added to a list that offered none	D3DMetal	12	4.0b2	26.3 and Preview	Fixed
Devil May Cry 5	RE Engine	Crashes when a skill preview video plays	Staged VC-1 codec. Nothing installed beside the game	D3DMetal	12	3.0 and 4.0b2	26.3 and Preview	Fixed
RESIDENT EVIL 2	RE Engine	Crashes when a video plays	The same staged VC-1 codec, unchanged	D3DMetal	12	3.0 and 4.0b2	26.3 and Preview	Fixed
RESIDENT EVIL 3	RE Engine	Crashes when a video plays	The same staged VC-1 codec, unchanged	D3DMetal	12	3.0 and 4.0b2	26.3 and Preview	Fixed
NINJA GAIDEN 4	Koei Tecmo, in-house	Says the VP9 codec is missing, then exits	Staged Matroska demuxer, and the MFT gate answered	D3DMetal	12	3.0 only -- 4.0b2 stalls it	26.3 only -- Preview stalls before video	Fixed

Update the toolkit, then pick a CrossOver. Every fix here was written against Apple's Game Porting Toolkit 4.0b2, which is what CrossOver Preview ships and what CrossOver 26.3 does not -- 26.3 carries D3DMetal 3.0, and on 3.0 these patches do not find what they were written to find. So 26.3 is a perfectly good engine for all of this once its toolkit is replaced, and a poor one until then. The app does the replacing, and keeps the original beside it.

The exception is NINJA GAIDEN 4, which is the other way round: it runs on 3.0 and stalls on 4.0b2, before its first frame and for reasons inside the toolkit that nothing here can reach.

The GPTK column is the one that decides, and it is newer than this table. Apple's Game Porting Toolkit is what actually draws these games, and CrossOver ships it inside the bundle rather than as something you pick: 26.3 carries D3DMetal 3.0, Preview 27.0 carries 4.0b2 and uses it unless CX_GRAPHICS_BACKEND_VERSION says otherwise. So "this only works on Preview" has, for at least two titles here, meant "this needs the newer toolkit" and nothing about Wine at all.

The two rows in bold are where that stops being a footnote. NINJA GAIDEN 4 runs on 3.0 and stalls on 4.0b2. Life is Strange runs on 4.0b2 and crashes on 3.0. Opposite requirements, same machine, so there is no single toolkit that serves the whole table and no version of CrossOver that is simply "better". Both were measured by moving the toolkit under a fixed CrossOver, which is the only way to separate the two.

Everything else in the column is derived rather than freshly run: a title measured on 26.3 was measured on 3.0, and one measured on Preview was measured on 4.0b2. A cell naming one generation means the other was never tried, not that it fails.

Backend and DX are not preferences, they are requirements. Persona 5 Strikers, Nioh and Nioh 2 only work on DXMT: all three need a shared D3D9 surface handle, and DXMT implements sharing where D3DMetal has none to build on. Nioh 3, despite the name, belongs with the other group -- it is D3D12 on D3DMetal and never touches D3D9, and NieR Replicant is D3D11 on D3DMetal. The other nine run on D3DMetal with the D3D12 renderer, which is also what keeps PSO precompilation -- -dx11 dodges some of these faults and costs permanent shader-compilation stutter.

Which CrossOver, and what "Preview" means. Every measurement here was taken against CrossOver 26.3 and crossover-preview-arm64-20260821, and the CrossOver column says which of the two a title was measured on rather than which it might work on. Seventeen of the nineteen run on stable 26.3, which inverts where this project started: stable was the exception and is now the rule, and the toolkit -- not the engine -- turned out to be the axis that decides most of these titles.

The two exceptions point in opposite directions. The Kingdom Hearts pair has only ever been launched on Preview, so its rows record an absence rather than a result and nothing is claimed either way. NINJA GAIDEN 4 is the reverse: it runs on stock 26.3 and stalls on Preview, and what stalls it is the toolkit, which executes command lists concurrently with no lever to turn that off.

Two lessons paid for by rows that were wrong for a while, and are worth more than the statuses they corrected:

• A staged codec is built against one CrossOver and is not usable under another. Persona 5 Strikers was recorded as not working on 26.3 after a first attempt failed there; the codec simply had not been built for 26.3 yet.
• A fix that reports itself installed is not necessarily loaded. NieR Replicant was recorded as Preview-only because its 26.3 runs died at the first video. The bridge was never executing in those runs: the registry override it depends on had gone missing, and the installer answered installed from the files alone.

None of these games needs CrossOver patched, wherever the container can be opened. That was not true when this project started, and it is the single biggest thing that changed. The qualifier is the whole of what remains, and it is a container question rather than a codec one.

Both builds decode VP9 the same way, and for a long time what only Preview could do was open a WebM -- which was the whole of the difference. DYNASTY WARRIORS ships 355 .webm cutscenes and could not get as far as decoding on stable, while Mortal Shell 2 ships the same codec in .mp4, which both builds handle. The plugin-by-plugin comparison that conclusion rested on is in Findings, under The container, not the codec.

That gap is now closed, and it was a missing plugin rather than a missing engine. Neither build ships a Matroska demuxer; Preview reached WebM by another route. Staging libgstmatroska beside the decoder gives stable one too, and NINJA GAIDEN 4 is where it was measured -- it plays on stock 26.3, video and all, with nothing patched into CrossOver. What this means for the titles above has not been re-measured: their rows still say what each was measured on, and DYNASTY WARRIORS in particular deserves a fresh run on 26.3 before its row changes.

Three titles need a codec no CrossOver ships -- VC-1 for Persona 5 Strikers, WMV3 for Nioh and Nioh 2 -- and it is staged beside the game rather than patched into it. Nioh 3 needs none: its video is already NV12 by the time Media Foundation is asked for it.

None of these fixes decodes anything. The frames existed all along; they were being crashed on, mislabelled, or thrown away.

Every row is a title we deliberately took on, and every claim on the linked page comes from a measurement on an installed copy.

Each row links to a page with the findings and the fix for that title.

Pages

• Games — the table above, what each fix needs from CrossOver, and how a row gets added
• Diagnosing a new game — the tools, and what each one answers
• Findings — what the nine have in common: root causes, the vtable slots each hook takes, the carrier DLLs, the container-versus-codec comparison, the open defect on 26.3, and what was tried and did not work. These pages hold the per-title findings; that one holds what is shared.
• Running the scripts directly — for working from a clone: reproducing a fix by hand, or building a proxy for a title that has none. Nothing here is needed to use a release.