MiniMax H3 Just Got 2.6x Faster (New Veda Attention)

Attention Benchmarks for MiniMax H3 Video Generation

The video benchmarks three attention options using the same prompts, references, seed, 15-second duration, and 20-step quality setting:

• Standard attention: Uses full attention and has the longest render times.
• Sage attention: Retains full attention but uses a faster quantized kernel. It was generally close to Standard in speed and visual results.
• Veda sparse attention: Uses a distilled predictor to identify important parts of the attention map—roughly the top 10%—and skips much of the rest. Its sparsity settings are intended to help preserve important visual details.

Benchmark results

Test Standard Sage Veda Dialogue / confrontation 23.7 min 9.9 min 9.1 min Emotional close-up with voice 22.8 min 10.4 min 8.6 min Humming and singing 22 min 10.5 min 8.7 min Rooftop action 23.6 min 9.9 min 9.4 min Transition between keyframes 21.8 min 9.5 min 8.1 min

Memory use in the dialogue test was reported at about 15 GB for all three options. Across tests, Sage was typically more than twice as fast as Standard while remaining visually similar.

Veda was usually the fastest, reaching up to about 2.7 times faster in the keyframe test. However, it sometimes produced softer faces or details, altered scenery, or more motion blur. The action test showed the clearest potential drawback of sparse attention: smearing during fast movement. In the keyframe test, Veda switched to the final doorway later and changed its color, although the overall image quality remained good.

Recommendations

• Sage: Best general-purpose choice for faster renders with results close to Standard.
• Veda: Useful for dialogue, close-ups, and rapid prompt testing when speed matters more than maximum clarity.
• Standard or Sage: Preferable for fast action and final shots where detail and fidelity matter most.

Workflow and setup

The video demonstrates a ComfyUI workflow with optional switches for Standard, Sage, and Veda attention. It includes reference inputs for up to nine images, as well as video and audio inputs, and controls for resolution, duration, and workflow bypass or activation.

The test setup uses a MiniMax H3 Singularity model, a Qwen VL 32B text encoder, an int8 video VAE, and standard FP32 audio. The attention comparisons use no LoRA, keeping the quality test focused on the attention method.

The workflow also offers a draft mode using a special LoRA at three steps, intended for quickly checking prompts and references. Quality mode uses 20 steps. The video mentions a free or pay-what-you-want workflow and setup guide.

Terminology note: The subtitles sometimes render “Veda” as “Beta”; the comparisons and conclusions refer to Veda sparse attention.

Main speaker and source

The AI Brief — the video’s presenter, workflow demonstrator, and benchmark reviewer.