SGLang Diffusion benchmarked MiniMax-H3 video generation on eight NVIDIA H200 GPUs and found its dense inference path runs up to 1.95 times faster than Diffusers with zero loss in output quality, according to a technical post from the SGLang Diffusion team published this week. Stack two additional lossy techniques on top of that lossless baseline and the number climbs to 6.24 times, but that figure comes at a measured cost to visual fidelity that the lossless number does not carry.
The distinction matters more than the headline speedup. Whether a lab can serve AI-generated video profitably depends on inference cost per clip, not on how fast a demo renders once. A model that looks impressive in a launch video but costs too much per second to generate at scale stays a demo. SGLang’s benchmark is one of the first public accountings of where that cost actually goes on current hardware, and it separates the free win from the one you have to pay for.
The free win comes from fused kernels. Rather than run each step on its own, SGLang collapses a group of them into single GPU kernels: AdaLN modulation, gated residuals, the SwiGLU activation, and QK RMSNorm paired with 3D RoPE. In isolated microbenchmarks, individual fused kernels ran 2 to 12.16 times faster than their eager-mode equivalents. This is the layer that produces the 1.85 to 1.95 times lossless speedup: the same denoising math, computed with less memory traffic and fewer kernel launches. No frames change.
The paid win comes from two techniques that approximate rather than compute. Cache-DiT measures how much shifts from one denoising step to the next, and where that delta sits under a threshold it recycles the earlier result and drops the step entirely. SubBlock sparse attention, built on NVIDIA’s block-sparse attention kernel, drops the key blocks that contribute least to each attention calculation rather than scoring every block against every other. Both cut real compute. Both also change the output.
SGLang reports every accelerated configuration against SSIM, a similarity score measured against the lossless render. The fastest tested profile, SubBlock at 0.80 sparsity combined with Cache-DiT’s stride mode, delivers 5.06x to 5.72x speedup on text-to-video and 5.86x to 6.24x on the frame-conditioned FL2VA task, the source of the 6.24x headline figure. Its SSIM lands between 0.76 and 0.91 depending on task and duration, with the text-to-video case degrading more than frame-conditioned generation. For a quality-first setup, the team recommends Cache-DiT alone: up to 2.99x speedup at 0.90 to 0.92 SSIM, no sparse attention involved. A middle profile, SubBlock at 0.75 sparsity with Cache-DiT’s stride mode, delivers 4.90x to 5.93x at 0.79 to 0.90 SSIM.
None of the accelerated profiles are free lunches, and SGLang does not present them as such. The post explicitly frames sparsity as “the fraction of key blocks allowed to be dropped,” and shows SSIM falling as the sparsity setting rises from 0.75 to 0.80. The team also notes it has more lossy techniques, including quantization and progressive resolution, that were not part of this benchmark, meaning the tested envelope is a slice of what SGLang’s stack can eventually claim rather than a final ceiling.
For teams building on generative video, the operating decision is which SSIM band a product can tolerate before a user notices the difference. A background-generation pipeline that never puts output next to the source frame can likely run the aggressive 0.76 SSIM profile and pocket the full 6.24x. A product where users compare a generated clip against a reference image, the FL2VA case that starts a clip from an uploaded photo, needs closer scrutiny of the quality floor before deploying anything past the lossless or Cache-DiT-only tier.
Reported by the SGLang Diffusion team, with contributions from the Cache-DiT team, NVIDIA, and Ant Group, in a benchmark post published on the LMSYS blog on August 26, 2026.
