128 GB · 1× DGX Spark (GB10)

MOSS-Image-SFT on one DGX Spark (128 GB), second unit

one DGX Spark (128 GB), second unit

Model makerOpenMOSSX @Open_MOSS GitHub Hugging Face Website

ComfyUIINT8 ConvRot (Comfy-Org format)measured 20 Oct 2026independent check passed

Rank in the class

2 of 3 · overall 37.8 of 100, as of 20 Oct 2026

Speed 20.1 of 100 (25 %) · my ratings 43.8 of 100 (75 %). Speed is the mean of the two medians: up to 10 s per image it counts 100, above that 1000 divided by the seconds; my ratings are the mean of my 8 ratings, minus 5, times 20. All image setups on 128 GB.

Median per image · enhancer off49.33 s1024×1024, 12 images

Median per image · enhancer on50.11 s1024×1024, 12 images

Enhancer, per text7.7 smedian of 12 texts, own run

Memory peak27420 MiBsystem memory peak (MiB, unified memory, not comparable with VRAM), enhancer-off run

Enhancer off

12 images, enhancer off

Portraitmy rating 8.5 / 10 · enhancer off

Image I1, seed 42, Portrait, enhancer off — **seed 42** · 49.27 s full PNG

Image I1, seed 123, Portrait, enhancer off — **seed 123** · 49.29 s full PNG

Image I1, seed 456, Portrait, enhancer off — **seed 456** · 49.35 s full PNG

Fire breathermy rating 8 / 10 · enhancer off

Image I2, seed 42, Fire breather, enhancer off — **seed 42** · 50.06 s full PNG

Image I2, seed 123, Fire breather, enhancer off — **seed 123** · 49.34 s full PNG

Image I2, seed 456, Fire breather, enhancer off — **seed 456** · 49.53 s full PNG

Action shotmy rating 6.5 / 10 · enhancer off

Image I3, seed 42, Action shot, enhancer off — **seed 42** · 50.04 s full PNG

Image I3, seed 123, Action shot, enhancer off — **seed 123** · 49.31 s full PNG

Image I3, seed 456, Action shot, enhancer off — **seed 456** · 49.17 s full PNG

Alien forestmy rating 5.5 / 10 · enhancer off

Image I4, seed 42, Alien forest, enhancer off — **seed 42** · 49.96 s full PNG

Image I4, seed 123, Alien forest, enhancer off — **seed 123** · 49.29 s full PNG

Image I4, seed 456, Alien forest, enhancer off — **seed 456** · 49.20 s full PNG

Enhancer on

12 images, enhancer on

Portraitmy rating 8 / 10 · enhancer on

Image I1, seed 42, Portrait, enhancer on — **seed 42** · 49.85 s full PNG

Image I1, seed 123, Portrait, enhancer on — **seed 123** · 50.02 s full PNG

Image I1, seed 456, Portrait, enhancer on — **seed 456** · 50.08 s full PNG

Fire breathermy rating 8.5 / 10 · enhancer on

Image I2, seed 42, Fire breather, enhancer on — **seed 42** · 50.17 s full PNG

Image I2, seed 123, Fire breather, enhancer on — **seed 123** · 50.06 s full PNG

Image I2, seed 456, Fire breather, enhancer on — **seed 456** · 50.15 s full PNG

Action shotmy rating 6.5 / 10 · enhancer on

Image I3, seed 42, Action shot, enhancer on — **seed 42** · 50.03 s full PNG

Image I3, seed 123, Action shot, enhancer on — **seed 123** · 50.26 s full PNG

Image I3, seed 456, Action shot, enhancer on — **seed 456** · 49.17 s full PNG

Alien forestmy rating 6 / 10 · enhancer on

Image I4, seed 42, Alien forest, enhancer on — **seed 42** · 50.24 s full PNG

Image I4, seed 123, Alien forest, enhancer on — **seed 123** · 50.17 s full PNG

Image I4, seed 456, Alien forest, enhancer on — **seed 456** · 50.27 s full PNG

Honest notes

What to keep in mind

This setup ran on a second DGX Spark (same chip and software versions where documented); the two other models ran on the first unit.
Layer skip runs over all steps (as the README says); the ComfyUI template ends it at half. The text encoder is INT8 (the template uses w4a8).
Memory is the system peak of the unified memory, not comparable with the VRAM numbers of the other cards.

How it was run

Setup

Model: MOSS-Image-SFT
Weights: huggingface.co/Comfy-Org/MOSS-Image
Quantization: INT8 ConvRot (Comfy-Org format)
Runner: ComfyUI
Software: ComfyUI commit b1a247c9, torch 2.14.1+cu130, aarch64
Settings: 50 steps, CFG 3, euler / sgm_uniform, shift 5, negative prompt set, layer skip 22 on the negative branch, 1024x1024, batch 1, seeds 42 / 123 / 456
Enhancer on: Prompt upsampler of MOSS's own text encoder (temperature 0.7, seed = image seed)
Hardware: NVIDIA GB10 (DGX Spark), 128 GB unified memory

Exact command

Exact command and run record

# 1) start ComfyUI (b1a247c9), default memory management, one GPU visible
python main.py --listen 127.0.0.1 --port 8199 --disable-auto-launch
# 2) run the bench: enhancer texts first, then 1 warm-up, then 12 images enhancer off and 12 images enhancer on
python bench_bild.py moss --phase all
# the workflow graph used for every image is workflow_api.json in this folder

command.txt · workflow graph · all 4 prompts · times, enhancer off · times, enhancer on · All raw files

Prompts

The four house prompts

I1 · Portrait, I2 · Fire breather, I3 · Action shot, I4 · Alien forest. The full text of the four prompts is in prompts.json.

Credits

Also made this possible

Runner and INT8 weightsComfy Org (ComfyUI)GitHub Hugging Face Website

How to read this

What is like for like here

Every setup made the same 24 images: 4 house prompts (portrait, fire breather, action shot, alien forest; the full text is in prompts.json on each recipe page) × 3 seeds (42, 123, 456) × the prompt enhancer off and on, at 1024×1024, one warm-up image thrown away. The speed numbers are the medians of the 12 images of a mode. With the enhancer on, the model's own prompt rewriter wrote the text first; those texts were written once on the RTX 5090 box and reused on every box, so the enhancer-on images all start from the same text. The enhancer time is measured in its own run and is not in the image time.

The ComfyUI setups use INT8 ConvRot weights (Comfy Org format); the Mac uses mflux with its own 8-bit quantization, so the Mac recipe differs. The same seed gives other pixels on other hardware. I rate pictures I like to look at; it is one person's taste, so read the ratings as my ratings, not as a quality measurement. Image files are published with the metadata chunk removed, pixels identical; the hashes in the time files are those of the published files (png_sha256_original is the hash of the original). All images on this page are AI-generated.