dschaerer/Wan2.2-S2V-14B-GGUF

Wan2.2-S2V-14B GGUF — Noxalis Lab port

GGUF weights for Wan 2.2 S2V 14B (speech-to-video), for the NoxalisLab fork of stable-diffusion.cpp (branch s2v-port), which added S2V support to the C++ ecosystem.

• wan2.2-s2v-14b-q8_0.gguf — the S2V DiT, Q8_0 (norms/bias kept high precision). Verified numerically against the PyTorch reference: final output cosine 0.999996.
• wav2vec2-large-xlsr-53-f16.gguf — the speech encoder, f16 (norms/bias f32), weight_norm fused at conversion in float64. Per-layer NRMSE gates against the reference pass 33/33.

Also needed, from their existing hosts: umt5-xxl encoder (city96/umt5-xxl-encoder-gguf) and the Wan 2.1 VAE (Comfy-Org/Wan_2.1_ComfyUI_repackaged).

sd-cli -M vid_gen \
  --diffusion-model wan2.2-s2v-14b-q8_0.gguf \
  --audio-encoder wav2vec2-large-xlsr-53-f16.gguf \
  --t5xxl umt5-xxl-encoder-Q8_0.gguf \
  --vae wan_2.1_vae.safetensors \
  -p "a person speaking in front of a plain background" \
  --ref-audio voice.wav --init-img face.png \
  -W 480 -H 832 --video-frames 17 --fps 16 --steps 16 \
  --cfg-scale 4.5 --flow-shift 3.0 --sampling-method euler --diffusion-fa

Measured on a 128 GB Apple Silicon Mac (Metal): 1077 s for 17 frames at 16 steps, photorealistic. One segment (~1 s of audio) per run for now.

These weights power the Images & video view of Galactus Desktop (v0.1.21+).