Skip to content

FaceFX-ecosystem compatibility

OpenFaceFX is a clean-room, functional replacement for the FaceFX authoring pipeline. It does not read or write the proprietary FaceFX binary formats (.facefx, .fxa, .fxe, .ffxc) and is not affiliated with OC3 Entertainment or Speech Graphics.

This document is an honest survey (July 2026) of the FaceFX wrappers and tools that exist in the wild, and what "working with" each of them actually means.

The key fact about FaceFX wrappers

Every public "FaceFX wrapper" is a parallel generator, not a curve consumer: each takes audio + text and runs its own alignment to emit a game-specific lipsync artifact. None of them accept animation curves from another tool as input — so no lip-sync generator (ours or anyone's) can feed them directly. The practical integration surface is their output formats: pipelines built around a wrapper can swap in OpenFaceFX once OpenFaceFX writes the same artifact the wrapper would have produced.

Survey

Nukem9/FaceFXWrapper — the de-facto standard wrapper

  • Repo: https://github.com/Nukem9/FaceFXWrapper (MIT-labelled, last release v0.41, Dec 2021)
  • What it does: generates Bethesda .LIP lipsync files for Skyrim / Fallout without the Creation Kit. Used by xVASynth's lip_fuz plugin and the Mantella / Pantella AI-NPC pipelines.
  • Inputs: game type + FonixData.cdf (user-supplied, Bethesda-owned) + 16 kHz 16-bit mono WAV + dialogue text. Windows executable; embeds Creation Kit-derived code, so it is not itself clean-room.
  • Compatibility today: OpenFaceFX now writes Skyrim .lip directly from its phoneme segments (openfacefx.export_lip.write_lip, or -o out.lip on the naive/mfa commands) — an experimental clean-room writer. Path to full compatibility: in-game verification (volunteers welcome on issue #12). Format-research finding (Jul 2026, updated): the .fuz container and Skyrim's 16 speech-target names were already publicly verified; the .lip payload — a FaceFX facial-animation blob with no public byte-level spec — has now been reverse-engineered from four real samples (three mod-author placeholders plus one vanilla Creation-Kit asset). It is a frame-major curve grid; our codec (tools/lip_codec_research.py) re-serializes all four samples byte-identically and the writer's own output round-trips exactly. Shipped, but flagged experimental because the game has not yet loaded a written file and two facts stay unverifiable without the engine: the slot→morph assignment (numbered curve slots, not names) and the header u22 field (we copy the vanilla asset's value). Note the "12-byte version/size/flags header" is FaceFXWrapper's interface view — the on-disk payloads (including the .fuz-embedded vanilla asset) begin directly with the 24-byte FaceFX animation header. No Creation Kit code, FonixData.cdf, or game assets are used: format facts are derived by analysis and the writer is original code.

Calibrating the slot → mouth-target map (help wanted, ~20 min, no tools)

The one unresolved fact is which numbered curve slot drives which mouth morph — the payload routes by slot index, not name, so SKYRIM_SLOT_MAP in openfacefx.export_lip is a hypothesis (only its jaw guess at slot 22 has any evidence). You can resolve it in-game with nothing but a save and your eyes:

python -m openfacefx lip-calibrate --out calib

writes calib/slot_00.lip .. slot_32.lip (one per Skyrim grid slot) plus a README.txt. Each file sweeps exactly one slot 0→1→0 (~2 s) with everything else at rest. Then, for each file:

  1. Pick any voiced NPC line you can retrigger in-game.
  2. Substitute slot_NN.lip for that line's lip data (rename it in place, or repack the line's .fuz — openfacefx.bethesda.write_fuz does this).
  3. Play the line and note which part of the face moves (jaw open, lip closure, lip round, tongue, brow, none…). Record slot NN → <part>.
  4. Repeat across all 33 files, then post your slot→part table on issue #12.

That table is exactly what SKYRIM_SLOT_MAP needs; once confirmed, the experimental writer's mouth shapes become correct rather than provisional.

FaceFX/FaceFX-UE4 and FaceFX/FaceFX-UE5 — official Unreal plugins

  • Repos: https://github.com/FaceFX/FaceFX-UE4, https://github.com/FaceFX/FaceFX-UE5 (actively maintained; MIT for the plugin interface code only)
  • What they do: load compiled, actor-bound .ffxc / .facefx assets authored in FaceFX Studio. The FaceFX Runtime itself is not open source — it is an EULA-gated binary from facefx.com (a widely repeated claim that a FaceFX/Runtime source repo exists on GitHub is false).
  • Compatibility: fundamentally incompatible. Producing .ffxc requires the proprietary Runtime compiler, and compiled assets are locked to the actor they were built for. The supported route into Unreal is to bypass these plugins entirely: OpenFaceFX JSON → engine curve assets / morph-target tracks (the 15 channels are plain named float curves).

yokimklein/H3EK-FaceFXWrapper — Halo 3 / Reach editing kits

  • Repo: https://github.com/yokimklein/H3EK-FaceFXWrapper (C#, last release Jun 2022)
  • What it does: shims the Dragon Age-era FaceFX Studio build into the Halo editing kits and converts its LTF output into Halo's binary .FXX.
  • Compatibility: parallel audio+text generator for one game family; not a practical integration target.

Same-viseme-set neighbours (not FaceFX, but 1:1 with our output)

  • Meta/Oculus OVRLipSync and radiatoryang/lipstick (Unity fork): use the exact 15-viseme set OpenFaceFX emits (sil PP FF TH DD kk CH SS nn RR aa E I O U). Lowest-friction interop; an OVRLipSyncSequence/AnimationClip exporter is on the roadmap.
  • hecomi/uLipSync (Unity, MIT): blendshape-curve driven; reachable via a Unity AnimationClip exporter.
  • DanielSWolf/rhubarb-lip-sync: different (Preston-Blair-style) mouth-shape model; interop needs a viseme remap table, not a format bridge.

Summary matrix

Tool What it is Consumes OpenFaceFX output today? Route to compatibility
Nukem9/FaceFXWrapper audio+text → Bethesda .LIP Experimental Skyrim .lip writer shipped (-o out.lip); re-encodes real samples byte-exact, in-game verification pending (#12). .fuz container + .lip header tools also shipped (openfacefx.bethesda) Verify in-game; slot→morph map + header u22 are documented assumptions
FaceFX-UE4 / UE5 plugins load compiled .ffxc/.facefx No None (proprietary compiler required) — drive UE curves directly instead
H3EK-FaceFXWrapper audio+text → Halo .FXX No Not a practical target
OVRLipSync / lipstick Oculus 15-viseme runtime (Unity) Yes, via .anim export (write_unity_anim, -o out.anim) Shipped; OVRLipSyncSequence .asset deliberately skipped (version-coupled)
uLipSync Unity blendshape lipsync Yes, via .anim export Shipped
rhubarb-lip-sync audio → 2D mouth shapes n/a (different viseme model) rhubarb retarget preset shipped (docs/retargeting.md)

Consuming OpenFaceFX output yourself

The JSON is deliberately trivial — a reference reader is ~15 lines:

import json

def load_track(path):
    d = json.load(open(path))
    assert d["format"] == "openfacefx.track" and d["version"] == 1
    # {"PP": [(t0, v0), (t1, v1), ...], ...}  — linearly interpolate between keys
    return {c["name"]: [tuple(k) for k in c["keys"]] for c in d["channels"]}

Channel names are blendshape names from the Oculus 15-viseme convention; values are weights in [0, 1]; keys are [time_seconds, value], sorted, sparse (RDP-reduced). Retargeting to another rig is a name-remap plus optional weight scale — see visemes.py (PHONEME_TO_VISEME, VISEMES).

Trademark note

FaceFX® is a registered trademark of OC3 Entertainment, Inc. (whose assets were acquired by Speech Graphics in September 2025). OpenFaceFX is an independent open-source project: not affiliated with, endorsed by, or connected to OC3 Entertainment or Speech Graphics, and it contains no code or data from FaceFX products. The name is used descriptively to indicate the category of tool it replaces.