Skip to content
READY 2026-11-05
Image generation overnight batch: 180 jobs in 5.6 hours (5 local models, 1024x1024)
Hardware and Config
- GPU: 2x NVIDIA RTX PRO 6000 Blackwell Max-Q 96 GB (per purchase invoice; we have not measured the hardware ourselves)
- CPU: AMD Threadripper PRO 9975WX
- OS: Windows 11
- Date: 2026-11-05 (batch window 05:31-11:08 UTC)
Software Versions
- Models: qwen-image-2512, qwen-image-edit-2511, z-image, moss-image-sft, flux2-klein-4b, flux3-klein-distilled
- Precision: bfloat16 (as recorded per job)
- Size: 1024x1024 (as recorded per job)
Raw Numbers
The full overnight batch: 180 jobs, 180 distinct job ids, wall time ~5.6 h (~32 images/hour).
| Model | Jobs | Gen median (s) | Load median (s) | Total median (s) |
|---|
| qwen-image-2512 | 68 | 64.8 | 25.0 | 90.4 |
| flux2-klein-4b | 36 | 1.5 | 0.0 | 2.0 |
| moss-image-sft | 32 | 50.2 | 21.3 | 71.8 |
| z-image | 24 | 67.0 | 0.0 | 67.4 |
| flux3-klein-distilled | 11 | 6.8 | 13.1 | 20.2 |
| qwen-image-edit-2511 | 9 | 120.7 | 19.7 | 140.9 |
What the Numbers Do and Do Not Show
- DO show: how many images the box pushes in a real overnight run, per model, with load time separated from generation time
- DO NOT show: image quality or prompt adherence - pure throughput
- NOTE: the Qwen median here is higher than IMGGEN-001's 42s because this batch includes the edit route (qwen-image-edit-2511) and longer prompts; IMGGEN-001's 52-job subset is the like-for-like speed comparison
AI Disclosure
This reconciliation was run and written by an AI agent (OpenCode Build for Danny)
on the Genesis box. All numbers are recomputed from the attached raw-numbers file.
Link Slots
- Raw numbers: docs/recipes/data/IMGGEN-002.json
- Source CSV: docs/recipes/raw/image-job-timings-20261106.csv (sha256 a7bcd3cd913530d8...)
- Rerun script: tools/recipes/rerun_imggen001.py (same CSV; add a --overnight filter for this exact window)
- Related posts: IMGGEN-001 (the 52-job recorded subset)
Status
READY