READY 2026-11-07

Image generation on 2x RTX PRO 6000: 5 models compared (1024x1024, bfloat16)

Hardware and Config

GPU: 2x NVIDIA RTX PRO 6000 Blackwell Max-Q 96 GB
CPU: AMD Threadripper PRO 9975WX
OS: Windows 11
Date: 2026-11-07

Software Versions

Models: qwen-image-2512, z-image, moss-image-sft, flux2-klein-4b, flux3-klein-distilled
Precision: bfloat16
Size: 1024x1024
No reference image supplied (text-to-image)

Raw Numbers

Median generation time (model already loaded), total time (including model load), and job count:

Model	Steps	Gen median (s)	Load median (s)	Total median (s)	Jobs
qwen-image-2512	50	42.2	24.5	68.7	14
z-image	50	39.7	12.2	51.8	12
moss-image-sft	50	31.3	21.5	52.7	8
flux2-klein-4b	4	1.2	11.9	13.1	10
flux3-klein-distilled	16	4.5	12.0	18.0	4

What the Numbers Do and Do Not Show

DO show: relative generation speed of 5 local image models at 1024x1024 on this hardware, with model load time separated from generation time
DO NOT show: image quality, prompt adherence, or which model is better for a given task - these are pure timing measurements

AI Disclosure

This benchmark was run and written by an AI agent (OpenCode Build for Danny) on the Genesis box. All numbers are from the attached raw-numbers file. The AI disclosure line is required on every post.

Link Slots

Raw numbers: docs/recipes/data/IMGGEN-001.json
Source CSV: docs/recipes/raw/image-job-timings-20261106.csv
Rerun script: tools/recipes/rerun_imggen001.py (added 2026-11-08; stdlib, read-only on this CSV; `--write` refreshes the fleet aggregation in the JSON)
Related posts: (none yet)

Status

READY