niloy629/avatar-weights

Model files for the avatar render service

These four files are the complete set required to run avatar_server.py (port 8991). Total: 1.4 GB.

file	size	sha256 (first 16)	purpose
generator_last.ckpt	1.3 GB	see SHA256SUMS	static IMTalker generator (audio → motion), nfe=3, cfg=1.24
renderer_narrow.pt	43 MB	see SHA256SUMS	distilled renderer (motion → 512² frames), ~6 ms/frame
ref.jpeg	28 KB	see SHA256SUMS	reference identity photo
ref_x.pt	1.6 KB	see SHA256SUMS	cached reference latent — removes the need for renderer.ckpt

Deliberately not included:

• renderer.ckpt (2.0 GB, full IMTalker renderer) — only needed if you want to rebuild ref_x.pt. The cached ref_x.pt makes it unnecessary.
• wav2vec2-base-960h (361 MB) — public. prepare.sh pulls it from facebook/wav2vec2-base-960h automatically.

Where these files live on 5090-06

generator_last.ckpt   /home/ubuntu/project/speech2avatar_imtalker_personaplex_8998/
                        checkpoints/live_winner/fullgen_static_2s_6400_resume/last.ckpt
renderer_narrow.pt    /home/ubuntu/project/imtalker_world_model/design_a/
                        checkpoints/renderer_narrow/best.pt
ref.jpeg              /home/ubuntu/project/speech2avatar_imtalker_personaplex_8998/
                        IMTalker/assets/3robert.jpeg
ref_x.pt              /home/ubuntu/project/avatar_8991/ref_x.pt

The convenience symlinks under /home/ubuntu/project/speech2face/weights/ point at those locations; the staged copies in /home/ubuntu/project/hf_upload/ are dereferenced copies ready to upload.

Upload

pip install -U huggingface_hub
export HF_TOKEN=hf_xxx
huggingface-cli upload <your-username>/avatar-weights /home/ubuntu/project/hf_upload . \
  --repo-type model

Then the consumer side is exactly:

export HF_TOKEN=hf_xxx
export HF_REPO=<your-username>/avatar-weights
./prepare.sh
./run.sh
curl -s -F "audio=@sample.wav" http://localhost:8991/render

Verified

The package was extracted to a clean directory, started with run.sh, and rendered a 9.89 s clip in 1.89 s (rtf 0.19, fps 25, 512²). All Python imports resolve inside the package — nothing reads the original /home/ubuntu/project tree.

Not yet exercised: the HF_TOKEN/HF_REPO download branch, because the files were not uploaded at test time (the test used MODELS_DIR links).