Master MiniMax H3: 10x Faster, Low VRAM & 2K Resolution | ComfyUI Workflow

Summary

SOTAI demonstrates ways to optimize the official MiniMax H3 text-to-video workflow in ComfyUI. The same techniques can also apply to image-to-video and reference-to-video setups. The tutorial focuses on faster generation, attention optimizations, and latent upscaling.

Acceleration LoRAs

Simply lowering the default 20 inference steps can significantly reduce video quality. SOTAI compares two acceleration LoRAs:

• The four-step version renders quickly but can make skin look shiny and plastic.
• The eight-step version produces more natural-looking results and is recommended as the better balance.

Lowering the LoRA strength was not a reliable fix and sometimes introduced motion artifacts.

Attention and caching

The workflow uses Sage Attention for the first three blocks to help preserve structure, then Soul Attention (rendered as “Solo” in the subtitles) for later blocks to improve speed. A block-cache node further reduces repeated computation.

In SOTAI’s test, combining these techniques made a 10-second clip nearly twice as fast as the standard setup, saving almost two minutes without a noticeable loss of quality or motion smoothness.

Soul Attention is experimental and may have installation or compatibility issues. Sage Attention combined with block caching is offered as a fallback.

Latent upscaling

Rather than generating directly at the target resolution, the workflow first creates a low-resolution base video. It then uses a MiniMax H3-specific latent-scaling node followed by a short, three-step noise-reduction pass.

In the example, a 0.3-megapixel base is scaled to 1.5 megapixels in 2 minutes 56 seconds, compared with about 30 minutes for a standard 1.5-megapixel render on an RTX 3090. SOTAI says the result remains usable, but warns that latent upscaling does not reduce VRAM use, since high-resolution latent frames still need to be processed in memory.

Workflow resources

SOTAI recommends updating ComfyUI to use its official MiniMax H3 template. The workflow and custom resources are collected in a Patreon post.

Main speakers and sources

• SOTAI — Presenter and source of the workflow demonstrations and performance comparisons.
• Kijai — Credited by the presenter with developing Soul Attention.
• Community sources — Reddit benchmarks and community use of the acceleration LoRA and custom scaling node are referenced. The scaler’s creator is not named in the subtitles.