Image and video generation on a DGX Spark

This installer serves Qwen-Image 2.1 on SGLang Diffusion: a 1024×1024 image takes 34.3 s at 40 steps and 7.5 s with its eight-step Turbo. MiniMax-H3 makes 4 s of 480P video with its sound in 126.2 to 127.9 s with its Turbo adapter, and 622.8 to 632.5 s at 50 steps without it. Both run on a 128 GB DGX Spark and are driven from the cockpit.

How fast is image generation on a DGX Spark?

Qwen-Image 2.1 on SGLang Diffusion, one image per request, measured on the reference box:

Size	Base, 40 steps	Turbo, 8 steps
512×512	7.8 s	1.8 s
768×768	19.5 s	4.3 s
1024×1024	34.3 s	7.5 s
1664×928	54.6 s	11.9 s
2048×2048	182.8 s	40.6 s

Time grows with the steps, and faster than the pixels: four times the pixels of a 1024×1024 image cost about 5.3 times the time. Memory peaked between 32.1 and 45.4 GiB across these sizes.

How does that compare with other setups?

On the same model with another stack, an owner running Qwen-Image 2.1 in bf16 with diffusers measured 28.3 s at 20 steps and 54.1 s at 40 steps for a 1024² image, and 4 min 16 s at 40 steps for 2048² (thread 384774).

In ComfyUI, another owner measured FLUX.1-dev at 32.61 s warm and 113.17 s cold for a 1024×1024 image, and FLUX.2-dev (fp8mixed) at 50.14 s warm (thread 368344). Those are other models, shown for scale, not a like-for-like test.

What does the image lane do?

It serves text to image, image editing and transparent (RGBA) images. Its Turbo, Qwen's eight-step distillation of the same model, is a second checkpoint of the lane: about four and a half times faster, and close to the base on photographs, portraits, illustrations and short text. A plain install puts the lane on the box, and its first Load builds its runtime and fetches its checkpoint: 46 GB (34 GB of checkpoint, 12 GB of runtime and its build). The Turbo fetches 32.5 GB at its first switch.

How fast is video generation?

MiniMax-H3, served with the SGLang cookbook's recipe, makes text to video with the sound made with the picture, and takes a first and a last frame to start and end from. For 4 s at 480P, on three prompts with the same seed:

	Base, 50 steps	Turbo adapter, 8 steps
End to end	622.8 to 632.5 s	126.2 to 127.9 s
Of which denoising	598.6 to 599.3 s	102.5 to 104.1 s
Of which video decode	22.8 to 23.0 s	22.8 to 23.0 s

Its first Load needs about 174 GB of free disk: 145 GB of checkpoint and Turbo adapter, 12 GB of runtime and its build, 17 GB for the runtime's own caches.

How do I use them?

From the cockpit: on its Lanes view, pick Qwen-Image 2.1 or MiniMax-H3 and press Load, one action that switches the boot, stops the lane that serves and starts this one. Then the Image and Video views drive them. The lanes listen on 127.0.0.1:30020 and 127.0.0.1:30022, on the box itself, and not through the API on port 30001.

From a terminal, in the repo folder on the box:

./prepare-lane.sh image   # the first time only, builds the lane
./switch-model.sh image   # switches the boot, then prints the stop and start commands to run

video works the same way, and image-turbo switches to the Turbo.

Can a 64 GB DGX Spark do it?

Not with this installer: both lanes need a 128 GB DGX Spark, and the installer refuses them on a 64 GB box before a byte is downloaded. What runs on 64 GB.

curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash