I’m getting intermittent hissing/buzzing pulses in H3 clips containing sustained sounds such as tank engines, wind or low thunder. It can sound like pulses of metallic scraping/loud hissing layered over otherwise good audio. The video renders correctly.
In the original output, four pulses are particularly noticeable during approximately the first 1.5 seconds of a 5 second clip. Clips without sustained ambience can render with clean audio.
Setup
- ComfyUI Desktop, ComfyUI v0.38.1; PyTorch 2.12.1+cu130
- Windows, RTX 4090 (24 GB), i9-14900KF, 64 GB RAM
- Model:
minimax_h3_ref2va_pruned_int8_convrot.safetensors - Audio VAE:
minimax_h3_audio_vae_fp32.safetensors - Original generation: 20 steps,
res_multistep/simple, Turbo/Lightning disabled - Five-second clip, using a Ref2VA template reduced to one reference image
Attempted fixes
Simplifying the sound description and removing specific engine terminology did not resolve it. I also tried two audio-only refinement passes using ComfyUI-H3-AudioRefine, retaining the original video frames. Both used Euler/simple, CFG 1 and video denoise 0:
| Test | Settings | Result |
|---|---|---|
| Refinement 1 | 6 steps; audio denoise 0.45 | Quieter overall, but hiss remained proportionately similar |
| Refinement 2 | 20 steps; audio denoise 1.0 | Fewer noticeable pulses, but artifact persisted; possibly a slight relative improvement |
The second refinement is approximately 1.4 dB quieter overall, so lower volume alone shouldn’t be mistaken for an improvement.
Cross-checks
H3 through Higgsfield produces the same type of artifact with the same prompt and reference image. Comparable tests with Seedance and Grok Imagine do not. It’s audible through both PC speakers and wired earphones, though much less obvious on a phone speaker.
I have the workflow JSON, original output and both refinement samples available, but the forum currently prevents new accounts from uploading attachments. I can provide them when permitted.
Has anyone encountered this specific pulsed artifact, or identified a sampler, model or audio-decoding setting that resolves it? I’d appreciate a targeted next test.