CompareToolsMethod @niklaslenz_ai

Home / 128 GB · Image

128 GB · 1× DGX Spark (GB10)

Qwen-Image 2.1 Turbo (NVFP4) on a DGX Spark

Twenty 1024×1024 images at 8 steps with the NVFP4 files of Qwen-Image-2.1-Turbo in ComfyUI on one DGX Spark, measured twice from a fresh start.

Model makerQwen (Alibaba)Hugging FaceX @Alibaba_Qwen
NVFP4 quantization (files used here)BennyDaBallHugging Face
RunnerComfyUIGitHub
ImageQwen-Image-2.1-Turbo (NVFP4 quant by BennyDaBall)measured 13 Oct 2026not rankedAI-generated outputs

Not ranked. Measured on this device only. Measured before the image test v2 protocol of 20 Oct 2026; shown separately and not ranked.

128 GB

DGX Spark 128 GB · NVFP4 · 1024×1024 · 8 steps

one NVIDIA DGX Spark (GB10, 128 GB unified memory)

Median time per image, run 17.09 s20 images queued at once, median of 19 intervals, first image not counted
Median time per image, run 27.10 ssecond run from a freshly started container
All 20 images, run 1141.61 sfrom first submit to last image
Server execution per image, median, run 16.98 sComfyUI "executed in" time
Warm-up image (not counted)14.32 sfirst image after container start, seed 999
Images identical between run 1 and run 220 of 20sha256 of every PNG, compare png_sha256 in both summary files
Outputs

Real outputs

Close-up portrait of an old sailor with a white beard and deep blue eyes, golden hour light, 85mm le
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1003, run 1. · 6.95 s Prompt: Close-up portrait of an old sailor with a white beard and deep blue eyes, golden hour light, 85mm lens, realistic skin. full size
A red fox jumping through fresh snow in a pine forest, frozen motion, crisp morning light, wildlife
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1004, run 1. · 6.98 s Prompt: A red fox jumping through fresh snow in a pine forest, frozen motion, crisp morning light, wildlife photograph. full size
An astronaut sitting on the edge of a crater on the Moon, Earth rising behind, ultra detailed, cinem
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1005, run 1. · 6.94 s Prompt: An astronaut sitting on the edge of a crater on the Moon, Earth rising behind, ultra detailed, cinematic. full size
A lighthouse on a cliff during a violent storm, giant waves crashing, lightning in the sky, oil pain
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1008, run 1. · 7.28 s Prompt: A lighthouse on a cliff during a violent storm, giant waves crashing, lightning in the sky, oil painting style. full size
A hot air balloon festival over Cappadocia at sunrise, dozens of colorful balloons, aerial photograp
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1010, run 1. Slightly blocky balloon patterns and fine stripes on the slope (probably quantization traits, not isolated). · 6.94 s Prompt: A hot air balloon festival over Cappadocia at sunrise, dozens of colorful balloons, aerial photograph. full size
A barista pouring latte art in a small cafe, warm window light, shallow depth of field, photograph.
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1014, run 1. · 6.95 s Prompt: A barista pouring latte art in a small cafe, warm window light, shallow depth of field, photograph. full size
A lion resting on a rock in the savanna at sunset, backlit mane, National Geographic style.
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1015, run 1. · 7.00 s Prompt: A lion resting on a rock in the savanna at sunset, backlit mane, National Geographic style. full size
A chalkboard cafe menu that reads "TODAY: OAT LATTE 3.50 / CROISSANT 2.20", morning light, photograp
AI-generated DGX Spark, NVFP4, 1024×1024, 8 steps, seed 1017, run 1. The board text matches the prompt; the board is cropped by the frame. · 6.99 s Prompt: A chalkboard cafe menu that reads "TODAY: OAT LATTE 3.50 / CROISSANT 2.20", morning light, photograph. full size
Setup

How it was run

  • Files: BennyDaBall/Qwen-Image-2.1-Turbo-NVFP4 at revision cb5a4e68 (NVFP4 transformer 5.69 GB, NVFP4 text encoder 9.33 GB, bf16 VAE 0.68 GB); sha256 of all three files equals the Hugging Face API values on every device
  • Engine: ComfyUI 0.39.0, commit 46493d82, official NGC PyTorch 26.07 container, comfy-kitchen 0.2.37; NVFP4 runs natively, no fallback message in the log
  • Flags: --disable-api-nodes --disable-all-custom-nodes; no SageAttention, no --cache-none
  • Workflow: 1024×1024, 8 steps, Euler, the saved Turbo sigma plan, CFG 1; one image per job
  • Method: 1 warm-up image (seed 999) dropped, then 20 prompts queued at once with seeds 1001 to 1020; time per image = difference of consecutive server completion timestamps, median of the 19 intervals after image 1; includes text encoding, VAE decode and PNG save
  • Each run started from a freshly started container (no cache hits: each log shows exactly 21 prompts); two runs per device
Notes

What to keep in mind

Measured on this device only, not ranked.

One GB10 was measured. The two runs show the spread between runs on this machine, not between two GB10 boxes.

The images come from run 1. Run 2 produced pixel-identical images. Images are the model's raw output, no retouching. Image 1 (fire breather, not shown) has a cropped dark figure at the right edge on this device; image 10 (balloons) shows slightly blocky patterns, probably a quantization trait that was not isolated. Visual impression only, no quality metric.

The machine ran with its standard GPU clock cap (value not published here). A dashboard container was running idle on the machine, without GPU use.

Only the transformer uses NVFP4 matrix multiplication; the text encoder runs on the full-precision path.

The ComfyUI log shows a harmless error block at start (the GUI workflow templates package is not installed) and a deprecation warning for --disable-api-nodes; neither affects the run.

The model weights come under the Qwen research license (Hugging Face license field "other", name qwen-research); read the model page before commercial use.

Exact command · Summary run 1 (GB10) · Per-image timeline run 1 (GB10) · Summary run 2 (GB10) · Per-image timeline run 2 (GB10) · ComfyUI workflow · sha256 of the model files · Hugging Face repo info (quant) · All raw files