drozbay/Wan2.2-S2V-14B-module

Wan2.2 S2V 14B — Module (ComfyUI)

The S2V-specific layers of Wan2.2-S2V-14B as a standalone module: 165 tensors, 4.01 GB bf16 (audio_injector, casual_audio_encoder, cond_encoder, frame_packer, trainable_cond_mask), bit-exact from the official weights.

Stack it onto any 16-channel Wan 14B T2V-style trunk (Wan2.1/2.2 T2V, finetunes like Bernini-R) and ComfyUI detects the result as an S2V model — audio lip-sync, reference image, and control video on whatever trunk you pick, no baked 32 GB checkpoints needed.

Usage

1. Put wan22_s2v_module_bf16.safetensors in ComfyUI/models/diffusion_models/.
2. Load with KJNodes DiffusionModelLoaderKJ: trunk as the model, the module via extra_state_dict (use DiffusionModelSelector).
3. Condition with the native WanSoundImageToVideo node (needs a wav2vec2 audio encoder in models/audio_encoders/).

Bernini trunks: to keep BerniniConditioning working in S2V mode, save wan_s2v_module_patch.py into ComfyUI/custom_nodes/ and insert the Wan S2V-Bernini Patch node after the loader. Chain WanSoundImageToVideo → BerniniConditioning for audio + in-context refs together.

Notes: the injectors were trained against the S2V trunk (closest to Wan2.2 low-noise / Wan2.1), so expect the weakest results on high-noise experts and heavy finetunes. I2V trunks (36-channel) are not compatible.

Credits

• Wan2.2 — Alibaba
• Approach reverse-engineered from rzgar/Bernini-R-S2V
• Extraction and node built with Claude (Claude Fable 5)