augustolucasg/audio2face-3d

Amplify Cooper morph deltas 25x

Strip skin from Cooper (render as static mesh)

Add Cooper MetaHuman avatar (34MB compact)

Tame mouthClose: reduce PP/FF/nn intensities, slower decay, hard cap 0.35

Hold-with-decay between CTC phoneme spikes → dense viseme curves

Switch to xlsr-53 multilingual phoneme model (supports Portuguese)

Lower CONF_THRESH 0.30->0.10, surface top phonemes in summary

duration 90->60s for tight quota

Argmax visemes + confidence threshold + zero NVIDIA mouthShrug/Roll/Press

Reduce GPU duration 120→90s (quota constraint)

Bypass Wav2Vec2PhonemeCTCTokenizer (espeak-ng dep) — use FeatureExtractor + raw vocab

Propagate viseme error via module-level global

Capture visemes import/inference error in summary

Pin transformers<5 (v5.x has breaking API changes)

Real lipsync: wav2vec2 phoneme model + IPA visemes + ARKit mapping

Bump Face Cap weight scale to 2.8x (match iPhone capture peak range)

Per-avatar weight scale (1.8x for Face Cap to compensate subtler morphs)

Camera: pull back 2.5x, scale zoom bounds to mesh size, prevent wheel leak

Add KTX2Loader + MeshoptDecoder for Face Cap textures

Add Face Cap photoreal avatar + dropdown selector + name aliasing

Vowel shapes (spectral centroid) + idle blinks

Tighter envelope: peak amp + 2-frame lead + attack/release + no jawOpen smoothing

Hybrid solver: envelope-driven jawOpen + suppress mouthClose/Shrug during open

LASSO coord descent + active-poses mask for proper sparse decomp

Silence-baseline subtraction to kill spurious mouthClose drift

Fix camera framing + stronger solver (L2=0.1, thresh=0.04, EMA smoothing)

Bundle brunette.glb; embed inline base64; switch to librosa.load for any audio fmt

ARKit blendshape solver + RPM avatar viewer

Add three.js point-cloud viewer with audio sync

Fix duration: keep 15 frames/call (no overlap) + trim tail to audio length

Fix ONNX input shapes (window=16000, emotion=[B,30,10], stride=8000)

Bump to gradio 5.x to escape 4.44.1 dependency hell

Patch gradio_client.utils for bool-schema bug

Disable show_api (sidesteps gradio 4.44.1 schema bug)

Pin huggingface_hub<1.0 — gradio 4.44 needs HfFolder

Fix hardware slug in YAML frontmatter

Initial Audio2Face-3D v3.0 ZeroGPU wrapper

initial commit