[Bug] Minimax h3 audio reference doesn't seem to be working #1882

subject_definitions
<Subject 1> is the character from <Picture 1>.
<Picture 1> is the reference image of the character used as the visual anchor for the target video.
<Audio 1> is the source audio providing the dialogue, timbre, and delivery for the target video.

retention_analysis
retention_analysis
<Subject 1> (appears in [Shot 1]): fully_preserved - The character's appearance, clothing, and identity are kept consistent with the reference.
<Picture 1> ([Shot 1] first frame): fully_preserved - The video begins exactly with the composition and subject pose shown in <Picture 1>.
<Audio 1>: full_copy

summary : [audio reuse + reference generation] The target video features <Subject 1> speaking the dialogue from <Audio 1> while maintaining the visual identity, clothing, and facial features established in <Picture 1>.

detailed_description
[Shot 1] For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. The scene opens with <Subject 1> in the exact pose and framing of <Picture 1>. The camera holds a static shot with a very slight push in. <Subject 1> (S1) looks directly into the lens and begins to speak, her lips and facial muscles moving in perfect synchronization with the dialogue from <Audio 1>

[INFO ] stable-diffusion.cpp:5973 - MiniMax-H3 Ref2VA: 1 image(s), 0 video(s), 1 audio clip(s)
[DEBUG] vae.hpp:142  - VAE Tile size: 16x16
[DEBUG] ggml_extend.hpp:884  - num tiles : 2, 2 
[DEBUG] ggml_extend.hpp:885  - optimal overlap : 0.750000, 0.750000 (targeting 0.250000)
[DEBUG] ggml_extend.hpp:886  - processing 4 tiles
[DEBUG] model_loader.cpp:1045 - loading 118/562 tensors from C:\sd-models\minimax_h3_video_vae_fp16.safetensors
  |##################################################| 118/118 - 960.80MB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 0.36s (read: 0.21s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_manager.cpp:409  - model manager prepared params backend buffer (344.04 MB, 118 tensors, RAM)
[DEBUG] model_manager.cpp:506  - model manager staged compute params (344.04 MB, 118 tensors) to ROCm0, taking 0.08s
[DEBUG] ggml_extend.hpp:2244 - minimax_h3_video_vae compute buffer size: 627.13 MB(VRAM)
  |==================================================| 4/4 - 4.88it/s
[DEBUG] vae.hpp:166  - computing vae encode graph completed, taking 2.21s
[DEBUG] model_loader.cpp:1045 - loading 135/917 tensors from C:\sd-models\minimax_h3_audio_vae_fp32.safetensors
  |##################################################| 135/135 - 910.11MB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 0.36s (read: 0.18s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_manager.cpp:409  - model manager prepared params backend buffer (329.46 MB, 135 tensors, RAM)
[DEBUG] model_manager.cpp:506  - model manager staged compute params (329.46 MB, 135 tensors) to ROCm0, taking 0.05s
[DEBUG] ggml_extend.hpp:2244 - minimax_h3_audio_vae compute buffer size: 417.04 MB(VRAM)
[INFO ] minimax_h3_audio_vae.hpp:470  - MiniMax-H3 audio VAE encode completed, taking 0.68s