light-avatar weights

Shared weights for bithuman-product/light-avatar — fast, geometry-exact, per-identity talking-head avatars distilled from the Ditto teacher.

This repo hosts the read-only inputs every light-avatar host needs: the Ditto teacher cfg pkl, the ONNX models (source-of-truth for any TRT-engine build), the CUDA GridSample3D plugin, and the identity-agnostic generic flow init used as a warm-start at build time.

It does not host per-identity .pt avatars (you build those yourself with python -m light_avatar.cli build) and it does not host pre-built TensorRT engines (those are GPU-arch-specific; build them locally from the ONNX via trtexec).

Fetch

pip install huggingface_hub
export LIGHT_AVATAR_HOME=~/light-avatar-home
python -m light_avatar.cli fetch-weights \
    --weights-repo stevegu1984/light-avatar \
    --profile build              # or: cpu-render | gpu-render | all

Profile subsets:

profile	size	what for
cpu-render	~250 MB	CPU rendering of pre-built avatars
gpu-render	~252 MB + your TRT engines	GPU rendering with locally-built TRT engines (needs the GridSample3D plugin)
build	~2.3 GB	full set needed to distill a per-identity avatar from scratch
all	~2.3 GB	same as build today

Idempotent + sha-verified: re-running just verifies the cache and skips downloads.

Layout

The repo mirrors the canonical LIGHT_AVATAR_HOME/teacher/ tree:

teacher/
├── ditto_cfg/
│   └── v0.4_hubert_cfg_trt.pkl
├── ditto_onnx/
│   ├── *.onnx                            # 12 models
│   └── libgrid_sample_3d_plugin.so       # CUDA plugin for warp_network engines
└── flow_student_generic.pt               # bf16, generic warm-start flow init

See docs/WEIGHTS_HOSTING.md in the main repo for self-hosting instructions, profile semantics, and pinning revisions for reproducible CI runs.

License

Inherits the upstream Ditto-TalkingHead license. The bundled identity-agnostic motion students (1-step LMDM + bitAudio HuBERT) are distilled from Ditto and ship in the main repo under the same license.