Fighting Game · Architecture

Audio, VFX, Cinematic & Signature Content Pipelines

A focused page within the Fighting Game Architecture documentation. The full map and every sibling page live in the Architecture hub.

9sections15 minread1diagram
← Back to the Architecture hubMarkdown source
On this page

This page covers the four presentation modules that turn a deterministic match into something an audience can feel: V2Audio (foley, voice, barks, dynamic music, spatialization, commentary), V2VFX (hit-sparks, blood, X-Ray, dismemberment, auras, stage destruction, weather), V2Cinematics (camera rigs, slow-mo, finishers, entrances, the replay viewer, MP4 export), and the cross-cutting per-fighter signature library and Character Presentation layer that bind those three together into a fighter's identity. Where the combat core has to be provably correct (see Combat System: GAS, Frame Data & Determinism), this layer has the opposite job: it is allowed to be as expensive and as hand-authored as the platform budget permits, provided it never desyncs a match. That single constraint shapes everything here. Every effect these modules emit is cosmetic, and the game-feel data they read is explicitly tagged rollback_budget_excluded and resim_cosmetic_suppressed in V2/balance/feel/hitstop_curves.csv — presentation is computed on the live frame and deliberately suppressed during rollback re-simulation, so a hit-spark or a slow-mo curve can never change the integer state the netcode hashes.

Architecturally, all three modules are built the same way and for the same reason. Each is a UGameInstanceSubsystem (UV2AudioSubsystem, UV2VFXSubsystem, UV2CinematicsSubsystem) that holds registered catalogs and exposes paired Resolve…/Queue… entry points, while the actual decision logic lives in a sibling …BlueprintLibrary of pure, static functions the subsystem delegates to. That split is not ceremony: it lets every resolution rule be unit-tested without a live world, and it keeps the math deterministic and side-effect-free so the same call produces the same answer on every machine. The source of truth is the Unreal C++ under V2/ue/Source/V2Audio, V2/ue/Source/V2VFX, and V2/ue/Source/V2Cinematics, the balance data under V2/balance/{dialogue,feel,cinema,audio,fighters}, and the compiled V2/ue/Plugins/V2AICommentary plugin. This page is part of the "Modes, Presentation & Interface" set; the section hub is ../V2_ARCHITECTURE.md.

What ships, honestly#

Real and tested (deterministic C++ that runs). The resolution logic in all three subsystems is genuine and heavily exercised: there are 26 audio, 9 VFX, and 37 cinematics dedicated *.spec.cpp suites under V2/ue/Source/V2Tests/Private/{Audio,VFX,Cinematics}/, plus module-level Audio.Module.spec.cpp, VFX.Module.spec.cpp, and Cinematics.Module.spec.cpp in V2/ue/Source/V2Tests/Automation/. These are not shape-only tests — they assert specific computed values. KOLastHitSlowMo.spec.cpp asserts that the Mortal Kombat Fatal-Blow curve reaches EvaluateDilationAt(0.08f) == 0.10f while the Tekken curve stays at 0.68f (readable), and that KO last-hit slow-mo is selectable by exactly four launch rulesets. The implementations behind those assertions are real algorithms (a piecewise-linear keyframe lerp for slow-mo, a gore-tier min-clamp for damage VFX, a fail-loud command router for middleware), not lookups of hardcoded results. The V2AICommentary plugin is compiled — V2/ue/Plugins/V2AICommentary/Binaries/Linux/libUnrealEditor-V2AICommentary.so exists.

A real seam, not a real integration (audio middleware). The Wwise/FMOD "SDK integration" surface is honest about being a seam. InitializeWwiseSoundEngine and friends build a request struct and call Plan.ResolveCommand(...) (V2/ue/Source/V2Audio/Private/V2AudioBlueprintLibrary.cpp:4105), which validates a plan and either routes the command or rejects it with a tag like Audio.WwiseSDK.Rejected.NativeFallbackRequired (V2/ue/Source/V2Audio/Private/V2AudioTypes.cpp:1055). No actual Wwise or FMOD runtime is linked: V2Audio.Build.cs depends only on Core, CoreUObject, Engine, and GameplayTags. The native runtime is MetaSounds + Audio Modulation (EV2AudioRuntime::MetaSoundsAudioModulation is the default), and the middleware layer is a deterministic plan that records what would be issued and falls back loudly when the SDK is absent — exactly the fail-loud pattern, not a faked success.

Specified and validated, but art-pending. None of these modules ship the authored binaries the monolith's Content/<…>.uasset paths imply. The only authored gameplay .uasset in V2/ue/Content/ are the eight DA_RegionVariant_* data assets (plus an editor-only V2FrameDataInspector.uasset); there are zero camera-rig, Sequencer-preset, KO-curve, or Photo-Mode assets. Camera rigs, slow-mo curves, and Sequencer presets are modeled as USTRUCT catalogs registered at runtime, not as Content/Cameras/… or Content/Curves/… binaries. Voice-over is text + path placeholders: every locale_x_text column in V2/balance/dialogue/<fighter>.csv reads TODO, and the vo_file_path entries point at /Game/V2/Audio/Voice/… assets that do not exist yet. The per-fighter signature checklists say so explicitly — see the bolded note in V2/balance/fighters/Bishop/signature-checklist.md: a checked item "attests that the launch-roster signature catalog … defines and validates the requirement — it does not attest that any game asset has been authored." Treat this whole page as logic shipped, content pending.

Provider-gated. MP4 export's NIL/likeness check (@themis/likeness, forthcoming) and C2PA provenance stamping, licensed-streaming soundtrack rights, and the quarterly ADR recording pipeline are external dependencies; the code carries the gates and manifests, but the rights themselves are out-of-repo.

The shared pattern: register, resolve, queue, snapshot#

Every subsystem in this set follows one lifecycle, and understanding it once explains all three:

  1. Register — a RegisterXCatalog(const FX& Catalog, FString& OutFailureReason) call validates a catalog struct and stores it. Registration is fail-loud: an invalid catalog returns false with a human-readable reason and is not stored.
  2. Resolve — a BlueprintPure ResolveX(Request) const runs the deterministic decision (which template, which curve, which mix) against the registered catalog and returns a resolved struct carrying a bAccepted flag and, on rejection, a ReasonTag.
  3. Queue — a BlueprintCallable QueueX(Request) calls ResolveX, checks IsValidResult(...), and only then appends to a pending list; a rejected result is dropped with its reason rather than silently spawned. For example UV2VFXSubsystem::QueueAuraEffectVFX queues only when Result.IsValidResult(&FailureReason) && Result.bAccepted (V2/ue/Source/V2VFX/Private/V2VFXSubsystem.cpp).
  4. Snapshot — CaptureRuntimeSnapshot() returns a flat struct of counts and "last resolved" fields (FV2AudioRuntimeSnapshot, FV2VFXRuntimeSnapshot, FV2CinematicsRuntimeSnapshot). This is the testable, introspectable surface the automation specs read back to confirm a registration or resolution actually happened.
flowchart LR Evt[Live match / replay event] --> Audio[UV2AudioSubsystem] Evt --> VFX[UV2VFXSubsystem] Evt --> Cine[UV2CinematicsSubsystem] Audio -- Resolve --> ABL[V2AudioBlueprintLibrary · pure] VFX -- Resolve --> VBL[V2VFXBlueprintLibrary · pure] Cine -- Resolve --> CBL[V2CinematicsBlueprintLibrary · pure] ABL -- accepted --> AQ[Queue + Snapshot] VBL -- accepted, gore-clamped --> VQ[Queue + Snapshot] CBL -- accepted --> CQ[Queue + Snapshot] GorePolicy[(Gore policy ceiling)] -. clamps .-> VFX NIL[(NIL / likeness gate)] -. blocks export .-> Cine Native[(Native-fallback seam)] -. routes .-> Audio AQ -- cosmetic, rollback-inert --> Frame[Presented frame] VQ --> Frame CQ --> Frame CQ --> Export[Replay viewer / MP4 export]

Audio pipeline (V2Audio)#

V2Audio is by far the largest of the three (V2AudioTypes.cpp alone is ~15.4k lines; V2AudioBlueprintLibrary.cpp ~7.5k), because it carries the full adaptive-music brain plus the middleware abstraction. The runtime baseline is MetaSounds + Audio Modulation (EV2AudioRuntime), and the FV2AudioRuntimeSnapshot in V2/ue/Source/V2Audio/Public/V2AudioSubsystem.h exposes hundreds of introspection fields covering each subsystem below.

  • Foley, voice, and barks. RegisterFighterFoleyCatalog + ResolveFighterFoleyLayer drive per-fighter footstep/cloth/weapon layers (EV2FoleyLayerKind); RegisterVoiceBank / QueueVoiceLine handle the per-fighter voice bank; and RegisterVoiceBarkCatalog / ResolveVoiceBarkEvent implement the priority + cooldown bark router on gameplay events (finisher and KO barks win all races).
  • Dynamic music. This is the richest part. Music intensity is a five-stage vertical stack — EV2MusicIntensity is Silence → Ambient → Tension → Climax → KOSting — driven by a real adaptive-music state machine (EV2AdaptiveMusicState: Exploration, Tension, Combat, Victory, Defeat, Menu). On top of that sit horizontal resequencing (verse/chorus/bridge via EV2HorizontalMusicSectionRole), vertical layering across seven instrument layers (EV2VerticalMusicInstrumentLayer: AmbientPad, Bass, Drums, Strings, Brass, Melody, Choir), intensity curves, one-shot stingers, sidechain-compression ducking under dialogue/SFX, beat-synchronized transitions snapped to beat/bar/phrase boundaries (EV2MusicSyncBoundary), tempo/BPM detection, and a Def Jam DJ-remix mixer. EvaluateMusicIntensityCurve updates the subsystem's CurrentMusicIntensity, CurrentAdaptiveMusicState, and active layer set only when the resolved result is accepted (V2/ue/Source/V2Audio/Private/V2AudioSubsystem.cpp).
  • Middleware abstraction (the seam). A unified provider layer (EV2AudioMiddlewareProvider) can route spatial sources, listeners, environment/reverb zones, occlusion raycasts, soundbank streaming, profiling, gameplay-event triggers, project generation, hot reload, fallback, and per-platform cook — for Wwise, FMOD, or a custom backend. As covered above, this resolves and validates commands but does not bind a real SDK; it routes to the native MetaSounds fallback when the middleware is absent.
  • Custom soundtrack, streamer-safe, and licensing. RegisterCustomSoundtrackPolicy / EvaluateCustomSoundtrackSource (EV2CustomSoundtrackSource, EV2MusicLicenseClass) gate user-imported tracks per platform TRC, and a streamer-safe catalog can auto-swap DMCA-risky tracks. Per-track license flags decide what may be uploaded to community galleries — the same per-import gate the entrance music slot enforces (below).
  • Spatial audio, commentary, crowd. Atmos / Tempest3D runtimes with a binaural HRTF fallback; a three-voice commentary stack with barge-in policy (EV2CommentaryBargeInPolicy) and ducking; per-fighter crowd chants, per-stage crowd noise, stamina-driven vocal modulation, a tournament "hush" mix, and announcer combo-callout tiers with language fallback. DualSense speaker routing and a microphone-consent gate (passive detection only; recording requires explicit consent) round it out.

VFX pipeline (V2VFX)#

UV2VFXSubsystem (V2/ue/Source/V2VFX/Public/V2VFXSubsystem.h) registers nine catalogs and exposes the same paired Resolve…/Queue… surface. The taxonomy is concrete and matches the design intent exactly:

  • Hit-sparks. EV2VFXStyle is the eight-spark library from the spec — Clean, Dust, Metal, Electric, Fire, Ice, Spirit, Blood — resolved by ResolveHitSpark to a FV2NiagaraTemplateSpec.
  • Gore, with a policy ceiling. EV2GoreTier is Off, Mild, Standard, Extreme. The important edge case is that the active gore policy clamps every request: ResolveCharacterDamageStateVFX min-clamps the requested tier to ActiveGorePolicy.GoreTier (V2/ue/Source/V2VFX/Private/V2VFXSubsystem.cpp:434), so a region or rating-board cap can never be exceeded by a caller asking for more. Blood is a full sub-pipeline: QueueBloodDecals, QueueGroundAndPoundBloodSplatter (EV2GroundAndPoundBloodPattern), and QueueBloodPool (EV2BloodPoolSpawnPolicy), with octagon-canvas absorption layers (EV2OctagonGroundTextureLayer) for the UFC ruleset.
  • X-Ray render pass. QueueXRaySkeletonPacket (and a cinematic variant, QueueXRayCinematicRenderPacket) builds packets across the X-Ray layers EV2XRayRenderLayer (TranslucentSkin, BoneStencil, BoneBreakFlash, OrganSilhouette, PierceThroughMask) with per-bone break types and animation states (EV2XRayBoneBreakType, EV2XRayBoneBreakAnimationState).
  • Dismemberment, hazards, trails, auras. ResolveDismembermentVFX is ruleset- and bone-gated (EV2CoreVFXRulesetScope restricts it to Mortal Kombat); stage hazards (EV2StageHazardVFX), weapon ribbon trails (EV2WeaponTrailProfileKind), and auras (EV2AuraEffectKind: DevilForm, SoulCharge, Heat, Burnout, the latter resolved against a normalized intensity) each have a resolve + queue pair.
  • Persistent damage and stages. Character damage layers (Sweat, Dirt, Bruise, Blood), a multi-tier damage-progression model with persistence rules, per-fighter signature damage VFX, stage destruction with reset policy (EV2StageDestructionResetRule), weather transitions (EV2StageWeatherKind), and atmospheric particles/god-rays. Every catalog has a b…CatalogValid snapshot flag so a cook can assert coverage before shipping.

Cinematic pipeline (V2Cinematics)#

UV2CinematicsSubsystem (V2/ue/Source/V2Cinematics/Public/V2CinematicsSubsystem.h) is the broadest by feature count, spanning camera, slow-mo, finishers, entrances, accessibility text, and the entire replay/export toolchain.

  • Per-ruleset camera rigs and slow-mo. Rigs key off EV2CinematicRuleset, which has nine members — the seven fighting rulesets (MortalKombat, StreetFighter, Tekken, WWE, UFC, SoulCalibur, DefJam) plus Racing and OpenWorld. Slow-mo is a real curve: EvaluateSlowMo looks up a FV2SlowMoCurveSpec by id and calls EvaluateDilationAt, a piecewise-linear interpolation over time-dilation keyframes (V2/ue/Source/V2Cinematics/Private/V2CinematicsTypes.cpp:356). ResolveKOLastHitSlowMo picks a per-ruleset KO style (EV2KOLastHitSlowMoStyle: Fatal-Blow strong, subtle, zoom-punch, broadcast-replay), which is what the balance file V2/balance/feel/slowmo_freeze_curves.csv encodes per ruleset (each row carries an accessibility_opt_out flag).
  • Finisher data assets. UV2FinisherData (V2/ue/Source/V2Cinematics/Public/V2FinisherData.h) is a UPrimaryDataAsset with a trigger condition (EV2CinematicTrigger, input motion/sequence, range window, required state tags), a soft cinematic sequence reference, a gore tier, a mode-eligibility bitflag (EV2FinisherModeEligibility: Fatality, Brutality, Friendship, Mercy, StageFatality, CriticalFinish, EnvironmentalFinisher), and eligible-ruleset tags. Its IsValidFinisherData validator is strict and domain-specific (V2/ue/Source/V2Cinematics/Private/V2FinisherData.cpp:78): a finisher must carry an id, a valid FinisherWindow trigger, a non-null cinematic ref, at least one mode bit, and an allowed ruleset tag (MK / Soul Calibur / Def Jam). Notably, finisher gore is restricted to Off, Mild, or Extreme — the intermediate Standard tier is deliberately excluded for finishers (IsSupportedFinisherGoreTier).
  • Entrance data assets. UV2EntranceData is the WWE/wrestling-grade entrance authoring model: ramp/curtain/aisle stage zones (EV2EntranceStageZoneKind), a pyro pattern with a safety radius and heat budget and a reduced-motion fallback (EV2EntrancePyroPatternKind), DMX-addressed lighting cues (EV2EntranceLightingCueKind), replay-safe camera cuts, pose-sequence montages, mic moments with captions and crowd response, a per-entrance music import slot (personal-use-only, with community-gallery upload blocked), and assignment scopes (Fighter, WWEUniverseShow, MyGMShow). A companion UV2EntranceEditorLibrary provides the editor-time validators (ValidateCameraCutSequence, ValidateAssignments, CanUseMusicImportSlotForEntrance).
  • Playback control: skip, resume, director cut, POV, commentary. Cinematics resolve skip granularity (EvaluateCinematicSkip), resume checkpoints (ResolveCinematicResumePoint), director-cut chapters, alternate POV, and an in-scene commentary track — each with a dedicated spec (CinematicSkipGranularity.spec.cpp, CinematicResumePoint.spec.cpp, DirectorCut.spec.cpp, AlternatePOV.spec.cpp).
  • Subtitles and captions. Speaker-tag styling, timing overrides, and cue captions (ResolveSubtitleSpeakerTag, ResolveSubtitleTimingOverride, ResolveCinematicCueCaption) make cinematic dialogue accessible.
  • Replay viewer and export. This is a full non-destructive editing surface layered over the deterministic input replay (the same stream V2Netcode produces — see Game Modes, Training & Replay): OpenReplayViewerSession, timeline edits, a free camera with orbit/entity tracking, a kill-cam (Start/Advance/CancelReplayKillCam), highlight detection (DetectReplayHighlights over EV2ReplayHighlightEventKind), and two export jobs — full QueueReplayExport and QueueReplayClipExport. Camera edits are keyframes (QueueReplayCameraKeyframe) layered on top of the original input, not edits to it. The MP4 export path is where the NIL/likeness gate and C2PA provenance stamping apply.

Character presentation & the per-fighter signature library#

The thing that makes a fighter feel theirs is not one module but a contract spanning all of them, expressed in two places.

The Character Presentation catalog (RegisterCharacterPresentationCatalog on the cinematics subsystem) carries the per-fighter presentation surface: bark triggers (EV2CharacterPresentationBarkTrigger: CounterHitLanded, ThrowLanded, ParryLanded, LowHP, ComboN, FinisherArmed, FinisherMissed, KO — the same gameplay events the audio bark router consumes), entrance styles (EV2CharacterPresentationEntranceStyle: WWEWalkout, TekkenStageArrival, MKPitSummon, SFHubArrival, DefJamClubStrut), nine ending-art styles (OilPainting, ComicBookPanel, AnimeCelShade, GothicWoodcut, HipHopGraffiti, PhotorealPortrait, UkiyoE, ArtNouveau, MangaInkWash), post-fight cutscene kinds (victory pose, UFC presser/mic interview, WWE mic moment), and unlock sources. The snapshot exposes bCharacterPresentationRosterReady, …MatchupsReady, …LoreReady, and …EntrancesReady so a cert pass can check roster coverage.

The Dialogue DB is real, per-fighter CSV. V2/balance/dialogue/<fighter>.csv (eight authored files: asha_storm, bishop_crowe, kai_voltage, lian_noor, mina_kade, nyx_river, rook_vale, sol_vega) carries the columns situation, opponent, ruleset, line_id, english_text, locale_x_text, vo_file_path, emotion, length_cap. Mirror matches use a dedicated mirror situation bank (e.g. Asha's Line.AshaStorm.Mirror.001, never the cross-matchup pool), and per-opponent lines override generic ones. The honest caveat: the locale_x_text column is uniformly TODO and the vo_file_path entries point at unrecorded assets — the text and routing ship; the localized VO does not.

The signature library is a per-fighter checklist at V2/balance/fighters/<Fighter>/signature-checklist.md (60 of them). Each tracks Section-72 requirements across animation, audio, VFX, stage, and HUD — eight unique impact stings, a signature crowd chant, stamina-driven breath layers, three dynamic theme variants, and so on. Crucially, the checklists are explicit that a checked box attests the catalog requirement is defined and validated (via UV2LaunchRosterLibrary::BuildDefaultFighterSignatureLibraryCatalog), not that any asset is authored; the "Asset authoring" section stays unchecked "until real assets land and pass the QA cert signature-library audit." This is the single clearest statement of where this whole layer sits: the contracts and resolution logic are real and tested; the art, audio, and VO they describe are the work that remains.

Director commentary, story theater & narrative expansion#

The "behind-the-scenes" layer has a real anchor and several spec-only extensions, and it is worth being precise about which is which.

Real and compiled: the V2AICommentary plugin (V2/ue/Plugins/V2AICommentary/). It defines validated catalogs — an Iris commentary cue contract, a post-match commentary spec, an auto-edited highlight reel spec (EV2AIHighlightShareFormat), a per-match analysis spec, an auto-clip library, a compliance spec, chat translation, and procedural dialogue — each built and validated through the BuildDefault… → Validate… → …Signature pattern (V2AICommentaryBlueprintLibrary). The cinematics narrative data is real too: V2/balance/cinema/cinematic-narrative-expansion.json carries four commentary track families (director, mocapBehindTheScenes, writer, internalMonologue), an ADR pipeline with quarterly session ids and per-event voice-line banks, and locale directors; BehindTheScenesCommentary.spec.cpp and DirectorCommentary.spec.cpp exercise the cinematics-side resolution.

Spec-only (referenced, not present): the monolith's V2/ue/Plugins/V2Mode_SideStory/, V2Mode_StoryTheater/, and V2Mode_CinematographyEditor plugins do not exist in the repo — the actual plugin set is V2AICommentary, V2AdaptiveAI, V2AssetLinter, V2Editor, and BellonaUnrealEditor. The Side-Story cinematic-plus-scripted-fights flow, the Story Theater browse/replay index with alt-language/subtitle/commentary toggles, the branch-tree visualizer, and the Ending Gallery are described in the architecture but are not yet code. The narrative-expansion data they would consume (commentary tracks, comics, ADR) is staged in V2/balance/cinema/ and the cinematics subsystem's FV2CinematicNarrativeExpansionCatalog, so the data contract exists ahead of the mode plugins.

Game feel, determinism, and failure modes#

The thread tying this page back to the combat core is the feel data under V2/balance/feel/. hitstop_curves.csv carries per-ruleset, per-reaction hitstop with attacker/defender/hitlag frame counts and — decisively — the columns rollback_budget_excluded and resim_cosmetic_suppressed set true on every row, so the feel layer is computed live and skipped on re-simulation. slowmo_freeze_curves.csv, camera_shake_curves.csv, and photo_mode_tone_curves.csv complete the set, each with an accessibility_opt_out lane. Accessibility is structural here, not bolted on: pyro and camera-cut definitions carry bReducedMotionFallback / bReducedMotionSafe, slow-mo rows carry accessibility_opt_out, and the microphone surface is consent-gated by construction.

Failure modes are uniformly fail-loud. A bad catalog fails registration with a reason string; a resolution that cannot be satisfied returns bAccepted = false and a ReasonTag (e.g. VFX.Aura.Rejected.InvalidResult, Audio.WwiseSDK.Rejected.NativeFallbackRequired) rather than spawning a degenerate effect; and the gore policy silently-but-safely clamps rather than honoring an over-budget request. Because every output is cosmetic and rollback-inert, the worst case for any bug in this layer is a missing or wrong effect on one client's screen — never a desync. That is the whole point of keeping presentation on the live frame and out of the hashed state.

For how these effects are triggered by combat, see Combat System: GAS, Frame Data & Determinism; for the montage notifies that fire foley and VFX, see Animation & Input Pipeline; for the replay stream the cinematics viewer edits, see Rollback Netcode & Tag-Team and Game Modes, Training & Replay.