[Question] Wan2.2-S2V-14B: audio_injector output projection weights near zero — is this expected?

[funnyfunny88]

Summary

When loading the public Wan2.2-S2V-14B checkpoint (HuggingFace Wan-AI/Wan2.2-S2V-14B), the audio injector output projection weights (audio_injector.injector.*.o.weight) are extremely small (std≈0.001), resulting in no measurable effect of audio conditioning on the model output.

Measured Data

All 144 audio_injector.* keys present (0 missing / 0 unexpected). Weight magnitudes:

Parameter	std	Assessment
q/k/v.weight	~0.008	Trained
adain.linear.weight	~0.008	Trained
o.weight	~0.001	Near zero_init
norm_q/norm_k	~0.000	Constant ~1.0

For comparison: blocks.0.self_attn.q.weight has std=0.015.

Controlled experiment (real audio vs zero audio, same input):

sigma	corr(flow_audio, flow_zero)	MAD
1.00	0.999133	0.039
0.50	0.999998	0.001
0.10	0.999998	0.001

100× boost of o.weight changes flow dramatically (std 1.49→0.81), confirming the audio pipeline (wav2vec2 → CausalAudioEncoder → AudioInjector) is functional — the bottleneck is specifically the o.weight magnitude.

Questions

1. Is o.weight std ≈ 0.001 the expected trained value for this checkpoint?
2. Are there additional audio fine-tuning weights that should be downloaded separately?
3. Has anyone else observed that the public checkpoint does not produce lip-synced video?

Environment

• Hardware: Ascend 910PremiumA (32GB) × 4
• Model: Wan-AI/Wan2.2-S2V-14B (downloaded Jul 2025)
• Framework: PyTorch 2.9.0 + torch_npu (fp16 inference)