FULL STORY

MiniMax H3 Tested: Quality Loss and Local Quantization Fails

Community tests of MiniMax H3 reveal significant image quality loss and instruction drift. Subsequent local deployments using quantized versions also suffered from severe facial distortions.

2026-08-03 ~ 2026-08-06 · 2 episodes · 11 posts

Episode 1 · MiniMax H3 Tested: Quality Loss, Quantization Issues and Prompt Drift (2026-08-03, 9 posts)

Community tests of the MiniMax H3 video model reveal multiple issues regarding image quality and prompt adherence. Users report significant detail loss in image-to-video (I2V) generation, making it inferior to text-to-video. Distant subjects and fine details (like eyes, teeth, and accessories) suffer from artifacts or severe corruption, which cannot be fixed by simply increasing the resolution to 1920px. Additionally, quantized deployment (U-net int8) results in universally distorted and blurry outputs, performing even worse than WAN2.2.

Confirmed

  • I2V experiences detail loss; while smooth in micro-movements, it struggles with fine details like accessories and teeth.
  • Quantized deployment severely degrades image quality, introducing noise and distortion.
  • The model fails to follow specific prompt instructions in certain scenarios, such as generating abnormal slow-motion or ignoring action requirements like switching hands to hold a cup.
  • There is a noticeable gap between direct generation using default workflows and the high-quality videos or official previews shared by the community.

Unconfirmed

  • It remains debated within the community whether these issues (like distant face corruption and I2V quality drops) are inherent model limitations or if they can be mitigated by adjusting workflows and parameters (e.g., using a facedetailer-like approach).

Why it matters

  • The discrepancy between the actual performance and promotional expectations directly impacts creators' workflow choices and deployment costs. Issues like the performance plummeting after quantization and distant subject corruption expose the model's limitations in local deployment and complex scenes, providing concrete pitfalls to avoid for future video model evaluations.

Episode 2 · MiniMax H3 Local Video Generation Faces Severe Facial Glitches (2026-08-06, 2 posts)

Developers reported severe facial distortions and eye artifacts when generating videos locally using the MiniMax H3 model. Despite its fast generation speed, issues persist even at higher resolutions, making its output quality inferior compared to models like Wan 2.2.