Autoresearch: image prompt optimization loop

[realityinspector]

Summary

Add a Karpathy-style autoresearch loop that autonomously optimizes Flash's image generation prompts by exploring the prompt mutation space and tracking quality vs cost.

What's in the branch

Branch autoresearch/flash/image-prompts adds:

• autoresearch/__init__.py — module docstring
• autoresearch/flash_autoresearch.py — the optimization loop

Mutation space

• Style keywords: photorealistic, cinematic, oil painting, watercolor, etc. (15 options)
• Lighting terms: golden hour, chiaroscuro, neon glow, etc. (14 options)
• Composition terms: rule of thirds, dutch angle, shallow DoF, etc. (13 options)
• Detail levels: minimal through ultra-8k (6 options)
• CFG guidance scale: 3.0–15.0 continuous
• Negative prompts: 6 options including none

Modes

• --dry-run: Synthetic CLIP scores derived from config hash (deterministic + small Gaussian noise). No API calls. Good for testing the loop mechanics.
• Live mode (not yet wired): Calls Flash API to generate images, scores with CLIP similarity.

Outputs

• results.jsonl — per-iteration records with config, CLIP score, cost, quality-per-dollar
• pareto.json — Pareto frontier of quality vs cost

Usage

PYTHONPATH=. python3 -m autoresearch.flash_autoresearch --dry-run --iterations 10

Next steps

• Wire live mode to Flash API endpoint
• Add CLIP scoring (or LLM judge) for generated images
• Integrate with SNAG-Bench for composite quality metric
• Run overnight optimization pass (~$50 for 50 iterations)