youssefhassan13/exp01-guidance-sweep

Exp 01 — CFG guidance sweep (SDXL + SD 3.5)

Pre-registered confirmatory experiment from the Operating System Hypothesis project: sweep classifier-free guidance on fixed prompt and seeds, compare convolutional UNet (SDXL) vs MMDiT (Stable Diffusion 3.5).

Prompt (fixed)

a watermelon, a glass half-filled with water, and a set of keys on a wooden table

Grid

• Guidance: 1.0, 1.5, 2.0, 3.0, 4.5, 6.0, 8.0, 11.0, 15.0
• Seeds: 42–51 (10 per setting)
• Unconditional baseline: empty prompt, same seeds
• Images per model: 100 PNGs (90 conditioned + 10 uncond)

Confirmatory result

Null on geometric structure M(g) on both architectures (|Spearman ρ| ≪ 0.3). Exploratory Suzuki-style rubric curves (veridicality, spontaneity, complexity) show a strong guidance dose–response on SDXL and a compressed profile on SD 3.5.

Layout

sdxl/
  g{guidance}_s{seed}.png    # conditioned outputs
  uncond_s{seed}.png           # empty-prompt baseline
  judgements.json                # blind VLM scores
  analysis_report.json           # confirmatory + exploratory stats
  figures/                       # M curve, rubric curves, contact sheets, GIFs
sd35/
  (same layout)
preregistration.json
manifest.json

Provenance

• Code / prereg: https://github.com/youssefhassan/operating-system-hypothesis/tree/main/experiments/01_image_test
• Git commit: 1f654b4b117d3672fd81abf9d8eeb2e61469c50d
• Exported: 2026-06-24T20:04:37.774154+00:00
• Judge: Claude Sonnet (blind, shuffled order)
• PNG count: {'png': 223, 'gif': 20, 'json': 6, 'other': 1}

Citation

If you use this dataset, please link the GitHub repo and note the pre-registration in preregistration.json. A formal paper citation will be added when available.

License

MIT — same as the parent repository. Generated images are research artifacts; underlying Stable Diffusion weights remain subject to their respective model licenses.