Problem or motivation

Problem

Tuning PaperBanana (VLM/image providers and models, refinement iterations, --optimize, --auto) means many manual runs and informal comparison of outputs/run_* folders. There is no single artifact that lists what was tried and how variants ranked.

Proposal

Add a paperbanana sweep command that:

• Takes one --input + --caption and sweeps over configurable axes (e.g. providers, models, iterations, optimize/auto).
• Writes each variant under sweep_<id>/variant_<nnn>/ and produces sweep_report.json with per-variant status, timing, and a ranked summary.
• Supports --dry-run to preview the variant matrix without API calls.

Proposed solution

• Add paperbanana sweep: one input + caption; axes as comma-separated lists (providers, models, iterations, optimize/auto); optional --max-variants and --dry-run.
• Add paperbanana/core/sweep.py: build the variant matrix (Cartesian product), parse values, rank successes, summarize.
• Per variant: merge into Settings, set output_dir to sweep_/variant_/, run PaperBananaPipeline.generate(), record timing + a simple score from the final critic suggestions; failures get an error field.
• Output: sweep_report.json with results and ranked list; dry-run writes planned variants only, no API calls.
• Tests: sweep helpers + CLI dry-run / bad axis input.

Area

Pipeline / agents

Alternatives considered

No response

Willingness to contribute

• I'd be willing to submit a PR for this feature