Skip to content

[Bug] Poor video and audio quality with Minimax H3 #1871

Description

@GlasslessPizza

Git commit

Using sd-master-b4f1fd6-bin-win-cuda12-x64.zip

Operating System & Version

Windows

GGML backends

CUDA

Command-line arguments used

sd-cli.exe -M vid_gen --diffusion-model minimax_h3_fl2va_pruned-Q4_K_M.gguf --vae minimax_h3_video_vae_fp16.safetensors --audio-vae minimax_h3_audio_vae_fp32.safetensors --llm qwen3vl_32b_minimax_h3-Q4_K_M.gguf --lora-model-dir lora --output video.avi -p "Fast paced car chase through a steampunk city, with the drivers loudly trashtalking each other in an obnoxious british smug tone, one of the drivers say 'You’re driving a glorified toaster with a spoiler. I’ve seen more aerodynamic shapes in a bowl of porridge'.lora:minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16:1" --cfg-scale 1.0 -v -W 1280 -H 768 --diffusion-fa --offload-to-cpu --rng cpu --fps 24 --video-frames 96 --steps 8 --seed -1 --batch-count 1

Steps to reproduce

Run the provided cmd on the binaries provided in the master-815-b4f1fd6 release.

What you expected to happen

Clean video quality without artifacts, clean audio quality with clearly audible voices and without compression noise and other audio artifacts.

What actually happened

Video quality is poor, with tons of pixel crawling and mpeg artifacts.
The audio quality is so poor I can't even make out what the characters are saying, I'd say it's comparable to extremely low bitrate mp3. Taking off the lora doesn't help.

Logs / error messages / stack trace

No response

Additional context / environment details

Model files gathered from the official documentation:
https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/minimax_h3.md

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions