TheMindExpansionNetwork/mind-drone-surreal-gallery

MIND DRONE — Surreal Gallery + Qwen-Image-2.1-Turbo Stress Test

Original AI-generated surrealist art gallery and model stress-test collection for the MIND DRONE project.

• 36 images total: 24 gallery pieces + 12 stress-test renders
• 36/36 verified, 0 failed
• Generated on a temporary NVIDIA RTX 4090 RunPod pod with ComfyUI 0.39.0 (master 08ff3c1)
• Executed model: Comfy-Org INT8 repack of Qwen-Image-2.1, using the official Qwen/Qwen-Image-2.1-Turbo 8-step sigma schedule
• License: Qwen Research License Agreement (inherited from the upstream model — see link above). Generated outputs carry the upstream model's terms; this dataset does not grant additional rights.

What's inside

• images/ — 36 original PNG renders (g01–g24 gallery, s01–s12 stress tests)
• jobs.json — full per-job ledger: prompt, seed, resolution, step count, schedule, prompt_id, queue wait, execution time, output size, SHA-256
• manifest.json — machine-readable provenance (hardware, runtime, model files, sampling config, totals)
• STRESS_REPORT.md — human-readable stress-test report with timings and findings
• contact-gallery.png / contact-stress.png — visual contact sheets
• run.py — the resumable batch runner (reproducible)
• mind-drone-surreal-gallery-v1.zip — packaged copy of everything above

Stress-test highlights

Test	Result
Native 2048×2048, official 8-step	✅ 40.7s, no OOM
Native 2048×2048, 4 / 12 / 20-step variants	✅ 23.4s / 56.6s / 90.3s
2752×1536 max-res widescreen	✅ 40.7s
Typography (MIND DRONE sign)	✅ exact text, readable
RGBA transparency	✅ real alpha channel (388k transparent px)
Very long prompt (~200 words)	✅ coherent

Download

pip install huggingface_hub
hf download TheMindExpansionNetwork/mind-drone-surreal-gallery --repo-type dataset --local-dir mind-drone-surreal-gallery
# or grab the single archive:
hf download TheMindExpansionNetwork/mind-drone-surreal-gallery --repo-type dataset mind-drone-surreal-gallery-v1.zip --local-dir .

Provenance & honesty notes

• All images are AI-generated by Qwen-Image-2.1 (Turbo schedule). They are original fictional artworks — no real people, brands, or locations.
• Timings in jobs.json are measured ComfyUI execution_success − execution_start values on this pod, not estimates.
• Visual QA was performed on contact sheets and key tiles. A few gallery tiles are intentionally surreal/ambiguous (e.g. g07, g18, g20, g23); these are noted in STRESS_REPORT.md rather than hidden.
• The temporary GPU pod was used only for generation and has been released; the images and metadata are preserved here and locally.