SoFlowRig v1.3-4 (DiT-B, MH 176-d)
Paper EMA weights for SoFlowRig v1.3-4: audio-driven MH Control Rig + global Chinese text emotion.
This is not v2.0 (TZ 117-d + mT5). Code: github.com/Uvvvvvvvv/SoFlowRig.
File
| File | Size | Contents |
|---|---|---|
soflowrig_ditb_400000_ema.pt |
~195 MB | ema_model only (no optimizer). finished_steps=400000, model_type=DiT-B, face_dim=176. |
Do not use the raw 781 MB training ckpt (model + ema + optimizer). inference_mha.py loads ema_model with strict=False.
wget https://huggingface.co/go49guy/SoFlowRig/resolve/main/soflowrig_ditb_400000_ema.pt
Architecture
- DiT-B: hidden 512, depth 8, 8 heads, FFN 2048
- face_dim=176: unidirectional 104 + bidirectional 72 (
with_rotate=false) - Audio:
facebook/wav2vec2-base-960h, aligned to 30 fps, DiT cross-attention - Text: frozen
BAAI/bge-base-zh-v1.5, CLS → 256-d, added to timestep conditionc - Training: SoFlow FM, velocity 75% + consistency 25%, 400k steps, batch 64, 1×H20
Main checkpoint 400000 was selected after a GPU A/B vs 150k: both produced valid 176-d JSON; 400k has lower train loss (Vel_MSE=0.033, Con_MSE=0.0027) and is the full schedule.
Inference
NFE=4, dual CFG (--cfg-scale audio, --text-cfg text). See the GitHub README. You still need local snapshots of wav2vec2-base-960h and bge-base-zh-v1.5 (not included here).
License
Apache-2.0 for this model card and the released EMA. Cite wav2vec 2.0 and BGE if you use the encoders.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support