MiniMax H3 audio sounds like a blown microphone (loud distortion)

[salty-fish-114514]

Hi! I've been playing with the new MiniMax H3 model and ran into an issue
with the audio output. Wanted to report it in case others hit the same thing.

What happens

The video comes out perfect, but the audio sounds like the microphone is
blown out — very loud with harsh crackling/popping, kind of like when you
scream into a cheap mic. The content itself is correct (the right sounds
happen at the right time), it just sounds distorted.

How to reproduce

• Official MiniMax H3 image-to-video template workflow, unchanged
• The smallest/pruned model set (nvfp4 qwen3vl, INT8 diffusion model, fp16 video VAE,
  fp32 audio VAE)
• Default sampler settings from the template
• Happens more often with loud/intense audio scenes (explosions, impacts,
  crowds). Changing the seed sometimes avoids it.

My setup

• ComfyUI 0.30.1 (Desktop)
• Windows
• AMD Radeon RX 7800 XT, 16 GB VRAM (ROCm 7.14)
• 64 GB RAM
• PyTorch 2.12.0+rocm7.14.0

Video generation works great otherwise — really impressed with the model,
just this one audio issue. Thanks!

[whitespace-rebel]

You could try using the euler sampler.
euler_ancestral, lcm, and dpmpp_3m_sde for example completely destroyed the audio during my test run.

[salty-fish-114514]

You could try using the euler sampler. euler_ancestral, lcm, and dpmpp_3m_sde for example completely destroyed the audio during my test run.

I will try

[whitespace-rebel]

After some other test runs, samplers i can recommend are:

• euler (sounds most natural/smooth)
• gradient_estimation (tighter/punchier sound, the low end is a bit "overdriven", but the high frequencies seem ok, smoother than dpmpp_2m and res_multistep)
• ipndm (but not ipndm_v)
• deis
• heun if you have time (sounds as smooth/natural as euler, but a bit higher quality).

Avoid any sampler with 'ancestral' or 'sde' for Minimax H3.

[Etroarl55]

I can't even get this functioning reasonably? I'm unable to get 0.3 megapixel 5 second output at 24fps at anything lower than an estimated 2 hours.

Wondering what you are doing differently or what your arguments are as I have the same system.

[salty-fish-114514]

I can't even get this functioning reasonably? I'm unable to get 0.3 megapixel 5 second output at 24fps at anything lower than an estimated 2 hours.

Wondering what you are doing differently or what your arguments are as I have the same system.

It is quite strange, I can finish a whole i2v run (0.8 megapixel 5 second 24fps) in 1.5 hour. I just uses the default Desktop deployment.

[Etroarl55]

I can't even get this functioning reasonably? I'm unable to get 0.3 megapixel 5 second output at 24fps at anything lower than an estimated 2 hours.
Wondering what you are doing differently or what your arguments are as I have the same system.

It is quite strange, I can finish a whole i2v run (0.8 megapixel 5 second 24fps) in 1.5 hour. I just uses the default Desktop deployment.

Ah I see, I thought you somehow got it working in a much faster way. I think the only difference is probably our startup arguements than. I don't think AMD will be able to get higher gen speeds until turbo lora or optimizations.

[salty-fish-114514]

After some other test runs, samplers i can recommend are:

• euler (sounds most natural/smooth)
• gradient_estimation (tighter/punchier sound, the low end is a bit "overdriven", but the high frequencies seem ok, smoother than dpmpp_2m and res_multistep)
• ipndm (but not ipndm_v)
• deis
• heun if you have time (sounds as smooth/natural as euler, but a bit higher quality).

Avoid any sampler with 'ancestral' or 'sde' for Minimax H3.

Thanks a lot! those sampler did help, but at the cost of increased generation time, its quite fair, LoL.

[whitespace-rebel]

@salty-fish-114514
This commit fixed the sound with the ancestral and sde samplers:
#15243

[salty-fish-114514]

@salty-fish-114514 This commit fixed the sound with the ancestral and sde samplers: #15243

Great!