distorted audio in MiniMax H3 video output on every environment and workflow tried (Desktop 0.31.0 & Portable 0.31.1)

### Prerequisites

- [x] I am running the latest version of ComfyUI
- [ ] I have custom nodes enabled

### What happened?

When generating a video with the MiniMax H3 (Ref2VA) workflow, the output audio track is consistently very poor / distorted ( harsh, noisy, unnatural sound), no matter which environment or workflow is used. The problem reproduces **100% of the time** across:

- Official clean ComfyUI **Desktop** app (v0.31.0, near-empty `custom_nodes`)
- Official **Portable** (nvidia) build (v0.31.1)
- The **official** H3 workflow (as shipped with the Desktop app)
- Third-party "accelerated" workflows (turbo LoRA, FirstBlockCache, etc.)

**Important:** In **all** of these runs the generation completes **successfully** — the queue finishes, the mp4 is written, and **neither the frontend (Web UI) nor the backend (console/server logs) reports any error, warning, or exception**. The poor audio is a *silent* quality problem, not a runtime failure.
##Important:## In ##all## of these runs the generation completes ##successfully## — the queue finishes, the mp4 is written, and ##neither the frontend (Web UI) nor the backend (console/server logs) reports any error, warning, or exception##. The poor audio is a #silent# quality problem, not a runtime failure.

A sample video is attached demonstrating the audio issue. **[ATTACH VIDEO]**

See this :
https://github.com/user-attachments/assets/778e8d68-7d44-4d8d-aaa6-f040dda26d87

##Steps to reproduce##

1. Install ComfyUI (Desktop 0.31.0 or Portable 0.31.1) with the official MiniMax H3 model set:
- `minimax_h3_ref2va_pruned_int8_convrot.safetensors` (UNET)
- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` (text encoder)
- `minimax_h3_video_vae_fp16.safetensors` + `minimax_h3_audio_vae_fp32.safetensors` (VAEs)
2. Load the official H3 workflow (or any H3 Ref2VA workflow).
3. Provide reference image(s) + reference audio (or leave audio empty / use built-in TTS via `` dialogue tags).
4. Run generation (any length up to ~15 s; also observed in multi-segment renders).
5. Play the output video: the audio track sounds distorted / unnatural.

##Expected behavior##

The output video should contain clean, intelligible audio matching the reference audio (or the spoken dialogue), with no harsh distortion, no buzzing/artifacts, and lip-sync aligned with the dialogue timing.

##Actual behavior##

The generated video plays with very poor audio quality, ##while every run reports successful completion with zero errors on both frontend and backend##. Local observations across attempts include:

- Harsh / distorted / noisy audio on the output track (the core symptom of this report).
- In reference-audio runs, the model sometimes reads out the reference audio content from the very start (0 s), regardless of the prompt timestamps (observed energy spikes in the first seconds, e.g. sustained speech at 65–74 dBFS during a 0–2 s segment that should be silent).
- Audio/lipsync misalignment in some runs.

##Environment##

#OS / Hardware#
- OS: Windows 11 Pro 26200 (build 10.0.26200.8875)
- GPU: NVIDIA GeForce RTX 5090 D v2 (24 GB)
- Driver: NVIDIA 610.88 (CUDA UMD 13.3)

#Environment A — Official ComfyUI Desktop (clean)#
- ComfyUI: ##0.31.0## (Desktop app, `D:\Comfy-Desktop`)
- Python: 3.13.12
- torch: ##2.12.1+cu130## (CUDA 13.0)
- custom_nodes: none (empty, pristine install) — reproduced with zero third-party nodes

#Environment B — ComfyUI Portable (nvidia build)#
- ComfyUI: ##0.31.1## (portable)
- Python: 3.13.14
- torch: ##2.10.0+cu130## (CUDA 13.0)
- Model files (identical across environments):
- `minimax3h/minimax_h3_ref2va_pruned_int8_convrot.safetensors` (~21 GB)
- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` (~15.7 GB)
- `minimax_h3_video_vae_fp16.safetensors` (~5.2 GB)
- `minimax_h3_audio_vae_fp32.safetensors` (~0.6 GB)
- Note: the Portable environment also had third-party H3-related nodes installed (e.g. FirstBlockCache, audio-T8, spectrum) — the issue reproduces identically in the clean Desktop environment, so it is not caused by those nodes.

##Workflows tried##

- Official H3 Ref2VA workflow (shipped with Desktop app): `MiniMaxH3ReferenceToVideo` + BasicGuider + BasicScheduler (simple/20 steps, res_multistep) + VAEDecodeAudio + CreateVideo + SaveVideo.
- Third-party "fast/accelerated" H3 workflows (turbo LoRA, FxAI frame calculator, etc.).

##Additional observations / related upstream threads##

- `Comfy-Org/ComfyUI#15470` — local H3 Ref2V no audio output + random lip movement.
- `Comfy-Org/ComfyUI#15326` — EasyCache audio stream regression.
- `Comfy-Org/ComfyUI#15390` (merged) — fix H3 audio corruption with EasyCache.
- `Comfy-Org/ComfyUI#15377` (merged) — fix full offload on minimax audio VAE.
- `MiniMax-AI/MiniMax-H3#17` — multi-speaker timbre crosstalk.
- Also observed that prompt-side declarations (`[reference generation + audio reuse]`, `: reference`) influence whether the reference audio is re-read, but do not fix the overall audio quality.

```

### Logs

```shell

```

### Additional Context

_No response_

Start by reproducing the issue with the official MiniMax H3 Ref2VA workflow and compare its VAEDecodeAudio/CreateVideo path across the clean Desktop and Portable environments. Review related issues 15470, 15326, 15390, and 15377 to determine whether the distortion is already addressed or belongs outside the frontend; done means identifying a reproducible root cause and producing clean, synchronized audio without runtime errors.

Written by the indexing model from the issue text.