32 GB · RTX 5090 32 GB

P-H3-Image against MOSS-Image and Qwen Image 2.1 on an RTX 5090

Six identical prompts, seed 42, one RTX 5090 32 GB. P-H3-Image makes 8.5 MP images in 15.7 s (w4a8) including its prompt upsampler; MOSS-Image-SFT makes 1 MP images in 8.6 s and Qwen Image 2.1 in 5.0 s.

Model maker (P-H3-Image, based on MiniMax-H3)Pruna AIHugging Face GitHub X @PrunaAI

Model maker (MOSS-Image-SFT)OpenMOSSHugging Face GitHub X @Open_MOSS

Model maker (Qwen Image 2.1)Qwen (Alibaba)X @Alibaba_Qwen

Maker of the base model MiniMax-H3 behind P-H3-ImageMiniMaxHugging Face

ComfyUI and the workflow templates used for Qwen Image 2.1 and MOSS-Image-SFTComfy-OrgGitHub GitHub

ImageP-H3-Image (w4a8, int8), MOSS-Image-SFT, Qwen Image 2.1measured 5 Nov 2026not rankedAI-generated outputs

Not ranked. Measured on the devices shown here only. Simpler protocol than the image test v2 of 20 Oct 2026: 6 prompts, seed 42, median of 6. Shown separately and not ranked.

32 GB

RTX 5090 32 GB · w4a8 · 2912×2912

P-H3-Image w4a8, one RTX 5090 32 GB

Median time per image (incl. prompt upsampler)15.7 s6 prompts, median of 6, seed 42, warm-up not counted

Sampler (diffusion steps)10.0 smedian of 6

Prompt upsampler3.0 smedian of 6

Slowest to fastest image14.4 to 17.4 s6 prompts

Image size2912×2912 (8.5 MP)width x height of the saved file

32 GB

RTX 5090 32 GB · int8 · 2912×2912

P-H3-Image int8, one RTX 5090 32 GB

Median time per image (incl. prompt upsampler)17.9 s6 prompts, median of 6, seed 42, warm-up not counted

Sampler (diffusion steps)10.4 smedian of 6

Prompt upsampler3.9 smedian of 6

Slowest to fastest image16.5 to 19.6 s6 prompts

Image size2912×2912 (8.5 MP)width x height of the saved file

32 GB

RTX 5090 32 GB · MOSS-Image-SFT · 1024×1024

MOSS-Image-SFT, one RTX 5090 32 GB

Median time per image (incl. prompt upsampler)8.6 s6 prompts, median of 6, seed 42, warm-up not counted

Sampler (diffusion steps)6.3 smedian of 6

Prompt upsampler1.9 smedian of 6

Slowest to fastest image8.3 to 9.9 s6 prompts

Image size1024×1024 (1.0 MP)width x height of the saved file

32 GB

RTX 5090 32 GB · Qwen Image 2.1 · 1024×1024

Qwen Image 2.1 with the template's short prompt, one RTX 5090 32 GB

Median time per image5.0 s6 prompts, median of 6, seed 42, warm-up not counted

Sampler (diffusion steps)4.6 smedian of 6

Prompt upsamplernone (template default)see notes

Slowest to fastest image4.9 to 5.3 s6 prompts

Image size1024×1024 (1.0 MP)width x height of the saved file

32 GB

RTX 5090 32 GB · Qwen Image 2.1 + P-H3 upsampled text · 1024×1024

Qwen Image 2.1 fed the text P-H3's upsampler wrote, one RTX 5090 32 GB

Median time per image5.2 s6 prompts, median of 6, seed 42, warm-up not counted

Sampler (diffusion steps)4.6 smedian of 6

Prompt upsamplernot in this time (text reused)see notes

Slowest to fastest image5.0 to 5.5 s6 prompts

Image size1024×1024 (1.0 MP)width x height of the saved file

Outputs

Real outputs

I2 · Poster with text

Retro space travel poster reading MARS EXPRESS with a white rocket over red dunes; small extra text MARS LINE and TICKETS ON BOARD that the prompt did not ask for — AI-generated P-H3-Image w4a8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 17.4 s Prompt: A retro-futurist travel poster for a fictional space line, bold lettering that reads "MARS EXPRESS", a rocket flying over red desert dunes, flat illustration. full size

Same poster idea as w4a8 with MARS EXPRESS, MARS LINE and TICKETS ON BOARD; different rocket angle and dunes — AI-generated P-H3-Image int8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 18.2 s Prompt: A retro-futurist travel poster for a fictional space line, bold lettering that reads "MARS EXPRESS", a rocket flying over red desert dunes, flat illustration. full size

Retro poster with MARS EXPRESS in italic letters and a rocket over red dunes, surrounded by a white border — AI-generated MOSS-Image-SFT (own prompt upsampler), one RTX 5090 32 GB, seed 42, 1024×1024. · 8.7 s Prompt: A retro-futurist travel poster for a fictional space line, bold lettering that reads "MARS EXPRESS", a rocket flying over red desert dunes, flat illustration. full size

Bold retro poster with MARS EXPRESS in large letters and an orange rocket over desert dunes — AI-generated Qwen Image 2.1 (short prompt), one RTX 5090 32 GB, seed 42, 1024×1024. · 5.0 s Prompt: A retro-futurist travel poster for a fictional space line, bold lettering that reads "MARS EXPRESS", a rocket flying over red desert dunes, flat illustration. full size

Dark blue poster with MARS EXPRESS, small MARS LINE and TICKETS ON BOARD text and a white rocket over red dunes — AI-generated Qwen Image 2.1 (P-H3 upsampled text), one RTX 5090 32 GB, seed 42, 1024×1024. The model received the text that P-H3's upsampler wrote for this request (see raw file). · 5.3 s Prompt: A retro-futurist travel poster for a fictional space line, bold lettering that reads "MARS EXPRESS", a rocket flying over red desert dunes, flat illustration. full size

I3 · Crowded market

Three or four shoppers at a produce stall on a wooden boardwalk with tomatoes and cucumbers; a close view, not a dense crowd — AI-generated P-H3-Image w4a8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 16.1 s Prompt: A crowded farmers market on a sunny morning seen from a slightly raised angle, many different people buying fruit and vegetables at colorful stalls. full size

Three or four shoppers at a produce stall on a wooden boardwalk; same scene as w4a8 with different details — AI-generated P-H3-Image int8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 16.7 s Prompt: A crowded farmers market on a sunny morning seen from a slightly raised angle, many different people buying fruit and vegetables at colorful stalls. full size

Dense, colorful crowd at fruit and vegetable stalls seen from a raised angle in bright sun — AI-generated MOSS-Image-SFT (own prompt upsampler), one RTX 5090 32 GB, seed 42, 1024×1024. · 8.6 s Prompt: A crowded farmers market on a sunny morning seen from a slightly raised angle, many different people buying fruit and vegetables at colorful stalls. full size

Wide view of a dense, dark-clothed crowd at a market, muted colors and soft detail, little sun — AI-generated Qwen Image 2.1 (short prompt), one RTX 5090 32 GB, seed 42, 1024×1024. · 5.1 s Prompt: A crowded farmers market on a sunny morning seen from a slightly raised angle, many different people buying fruit and vegetables at colorful stalls. full size

A few shoppers at a produce stall on a wooden boardwalk, close view, a woman in a straw hat on the right — AI-generated Qwen Image 2.1 (P-H3 upsampled text), one RTX 5090 32 GB, seed 42, 1024×1024. The model received the text that P-H3's upsampler wrote for this request (see raw file). · 5.3 s Prompt: A crowded farmers market on a sunny morning seen from a slightly raised angle, many different people buying fruit and vegetables at colorful stalls. full size

I5 · Fox in a terrarium

Red fox asleep on moss with mushrooms, large in the frame, with a glass pane behind it — AI-generated P-H3-Image w4a8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 15.2 s Prompt: A tiny red fox sleeping inside a glass terrarium full of moss and mushrooms, macro photo, warm light. full size

Red fox asleep on moss with mushrooms, large in the frame, glass pane behind; different pose than w4a8 — AI-generated P-H3-Image int8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 19.4 s Prompt: A tiny red fox sleeping inside a glass terrarium full of moss and mushrooms, macro photo, warm light. full size

Fox curled asleep on moss with red mushrooms inside a glass terrarium with a metal frame and a lamp above — AI-generated MOSS-Image-SFT (own prompt upsampler), one RTX 5090 32 GB, seed 42, 1024×1024. · 8.4 s Prompt: A tiny red fox sleeping inside a glass terrarium full of moss and mushrooms, macro photo, warm light. full size

Small fox asleep on moss with mushrooms inside a round glass bowl — AI-generated Qwen Image 2.1 (short prompt), one RTX 5090 32 GB, seed 42, 1024×1024. · 4.9 s Prompt: A tiny red fox sleeping inside a glass terrarium full of moss and mushrooms, macro photo, warm light. full size

Fox asleep on moss with mushrooms in front of a glass wall, close macro view — AI-generated Qwen Image 2.1 (P-H3 upsampled text), one RTX 5090 32 GB, seed 42, 1024×1024. The model received the text that P-H3's upsampler wrote for this request (see raw file). · 5.0 s Prompt: A tiny red fox sleeping inside a glass terrarium full of moss and mushrooms, macro photo, warm light. full size

I6 · Knight in a flooded cathedral

Oil painting of a knight in silver armor and red cape kneeling in a stone hall with light rays and stained glass; the knight is small in the frame and the floor is wet and reflective — AI-generated P-H3-Image w4a8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 14.4 s Prompt: A knight in ornate silver armor kneeling in a flooded cathedral, rays of light through stained glass windows, oil painting. full size

Oil painting of a kneeling knight in silver armor and red cape in a stone hall with light rays; the knight is small in the frame — AI-generated P-H3-Image int8, one RTX 5090 32 GB, seed 42, 2912×2912, shown here at 1280 px. · 16.5 s Prompt: A knight in ornate silver armor kneeling in a flooded cathedral, rays of light through stained glass windows, oil painting. full size

Oil painting of a knight in silver armor kneeling in shallow water inside a cathedral with stained glass windows — AI-generated MOSS-Image-SFT (own prompt upsampler), one RTX 5090 32 GB, seed 42, 1024×1024. · 9.9 s Prompt: A knight in ornate silver armor kneeling in a flooded cathedral, rays of light through stained glass windows, oil painting. full size

Oil painting of a knight in silver armor kneeling on a wet floor in a dim cathedral with light rays — AI-generated Qwen Image 2.1 (short prompt), one RTX 5090 32 GB, seed 42, 1024×1024. · 4.9 s Prompt: A knight in ornate silver armor kneeling in a flooded cathedral, rays of light through stained glass windows, oil painting. full size

Oil painting of a knight in silver armor kneeling on a reflective wet floor with colored stained-glass light — AI-generated Qwen Image 2.1 (P-H3 upsampled text), one RTX 5090 32 GB, seed 42, 1024×1024. The model received the text that P-H3's upsampler wrote for this request (see raw file). · 5.1 s Prompt: A knight in ornate silver armor kneeling in a flooded cathedral, rays of light through stained glass windows, oil painting. full size

Table

All setups on one RTX 5090 32 GB, same six prompts, seed 42, median of 6 per setup (listed in the order the setups were grouped, not ranked). Measured on this device only.

Setup	Image size	Steps	Prompt upsampler	Median time per image	Sampler	Upsampler	Range
P-H3-Image w4a8	2912×2912	8	bf16, in graph	15.7 s	10.0 s	3.0 s	14.4 to 17.4 s
P-H3-Image int8	2912×2912	8	bf16, in graph	17.9 s	10.4 s	3.9 s	16.5 to 19.6 s
MOSS-Image-SFT (own prompt upsampler)	1024×1024	25	MOSS own (template default)	8.6 s	6.3 s	1.9 s	8.3 to 9.9 s
MOSS-Image-SFT (short prompt)	1024×1024	25	none	6.3 s	5.9 s	-	6.3 to 6.4 s
MOSS-Image-SFT (P-H3 upsampled text)	1024×1024	25	P-H3 text, not timed here	6.6 s	6.1 s	-	6.5 to 6.7 s
Qwen Image 2.1 (short prompt)	1024×1024	25	none (template default)	5.0 s	4.6 s	-	4.9 to 5.3 s
Qwen Image 2.1 (P-H3 upsampled text)	1024×1024	25	P-H3 text, not timed here	5.2 s	4.6 s	-	5.0 to 5.5 s

Raw file

Setup

How it was run

Hardware: one RTX 5090 32 GB, ComfyUI 0.44.0 (master 7ad7adb3). Seed 42 on all six prompts. One warm-up image (seed 1) per setup is not counted.
P-H3-Image: Pruna's own ComfyUI workflow (P-H3-Image-Comfy), w4a8 and int8 weights, prompt upsampler (bf16) inside the workflow, 8 steps, euler, no CFG, 8 MP output (2912×2912).
MOSS-Image-SFT: Comfy template image_moss_image_t2i, int8 weights, 1024×1024, 25 steps, cfg 2.5, er_sde, the template's own prompt upsampler on.
Qwen Image 2.1: Comfy template image_qwen_image_2_1_t2i, int8 weights, 1024×1024, 25 steps, cfg 1, euler. Setup A uses the short prompt as the template does (no upsampler). Setup B feeds the text that P-H3's upsampler wrote for the same request.
Method: time per image = wall clock from submit to the saved file (includes the prompt upsampler where a setup has one). Median of 6 prompts; models were already loaded.

Notes

What to keep in mind

Measured on this device only (one RTX 5090 32 GB), not ranked. One run per prompt, six prompts, one seed: the range per setup is shown instead of an error margin.

The output sizes differ as the makers ship them: P-H3-Image renders 8.5 MP (2912×2912), MOSS-Image-SFT and Qwen Image 2.1 render 1.0 MP (1024×1024). That is 8.1 times the pixels, so the times are not per pixel.

Qwen B (and MOSS B in the table) reuse the text that P-H3's upsampler wrote. The 3.0 s that upsampler needed is not in their times.

Seen by eye only, no quality metric was run. P-H3's upsampler turned 'crowded' into 'several shoppers' (market scene) and added small poster text (MARS LINE, TICKETS ON BOARD) that the prompt did not ask for. The knight is small in the P-H3 frames. The fox in the P-H3 frames is large and the terrarium is barely visible.

MOSS-Image-SFT: the poster has a white border. Qwen Image 2.1 with the short prompt: the market scene is muted and soft, and the fox sits in a round glass bowl rather than a terrarium.

w4a8 and int8 draw the same request with different details (for example another knight pose), which fits the model card: 'w4a8 changes more of the details and sometimes the composition'. int8 took about 14 % longer than w4a8 here (17.9 s against 15.7 s).

Only 4 of the 6 prompts are shown as images (poster, market, fox, knight). The portrait and headphone prompts were measured and are in the timings but not shown.

The page shows P-H3-Image files resized from 2912×2912 to 1280 px; MOSS and Qwen files are shown at their native 1024×1024. The original PNG files are not published here.

License info (not blocking): P-H3-Image is under the MiniMax H3 Community License, which excludes the EU, the UK, South Korea and the US without a separate MiniMax license and asks for 'Powered by MiniMax H3' on products; read it before use.

Exact command · All timings (summary, medians per setup) · Prompts · MOSS-Image-SFT upsampler system prompt (Comfy template) · All raw files of this run (one job file per image with the full ComfyUI graph and upsampler text, images, commands) · All raw files