NVIDIA H3 Super Acceleration: MiniMax H3 Video up to 27.7x Faster

One of the standalone 1440p results on the NVIDIA page (10 s, 2560×1440): two-stage acceleration, no matched baseline was recorded for 1440p.

How it works

Instead of running many H3 denoising steps at full resolution, H3 Super Acceleration does most of the generation work in a short low-resolution draft and then restores high-resolution detail with a lightweight refinement pass. Each stage is one job, and the two stages run serially on one GB200:

Stage 1 generates the initial video at 896×512 in four denoising steps with the LightX2V MiniMax-H3 Turbo LoRA. Stage 2 upsamples the draft to the requested output resolution and runs a three-step Sol-Attn pass that restores high-resolution detail and consistency. The pipeline is not bit-exact with the SGLang baseline: it changes the sampling path, so differences can appear in detail, texture, motion or audio — the paired videos on the NVIDIA page let you judge that tradeoff directly.

Measured latency

End-to-end latency on one NVIDIA GB200, with the latest Sol-Super results using Sol-Attn for the complete Stage 2 service:

The two speedups come from different places. Stage 1 replaces the official H3 VAE decode (3.427 s) with the TAEH3 lightweight decoder (0.028 s), and Stage 2 replaces the dense LTX-2.5 Video VAE decoder (6.404 s in a matched 121-frame decode) with the Sol-Attn refiner and the LTX TAEHV final decode. The Stage 2 encoder is intentionally left unchanged: the refiner was trained on the original LTX-2.5 VAE latent distribution.

At 768p output the measured throughput works out to roughly 525 five-second videos per hour on one fully utilized GB200, versus 23.6 with the SGLang baseline.

Visual comparison

The NVIDIA page pairs each 768p and 1080p sample with a matching SGLang baseline run using the same generation inputs:

SGLang baseline	H3 Super Acceleration
A man and woman exchange a tense remark on a crowded Japanese street (1344×768, 5 s)	Same inputs, 22.2× faster

SGLang baseline	H3 Super Acceleration
Two martial artists face one another in a bamboo forest (1080p, 5 s)	Same inputs, accelerated

Two standalone 1440p results (10 s, 2560×1440) were also published without a matched baseline:

Availability

The pipeline is built from public components: the official MiniMax-H3 checkpoint, the LightX2V MiniMax-H3 Turbo four-step LoRA used for the draft, the LTX-2.5 pre-trained checkpoint, and the TAEH3 and LTX TAEHV lightweight decoders. It runs on NVIDIA Sol Engine with Sol-Attn on GB200 hardware, and is presented as a production reference with measured benchmarks rather than a ComfyUI node pack. At the time of writing there is no official ComfyUI workflow from NVIDIA for this pipeline; community ports of individual Sol Engine components (such as Sol-Attn patching) have started to appear.