# FLUX.2-klein Benchmarks

This folder holds durable benchmark inputs, outputs, and state for low-cost `flux2klein` testing.

## Current intent

- Keep overnight tests cheap and reviewable.
- Prefer one or two single-image smoke tests over broad matrices.
- Log enough context to reproduce a result without trawling temporary folders.

## Layout

- `canonical-inputs/`
  - durable reference inputs used for repeatable comparisons
- `runs/`
  - dated run batches
- `overnight-state.json`
  - handoff file between the `3am` optimisation pass and the `6am` review pass

## Batch contents

Each case folder should keep:

- `input-1.*`
- `resolved-prompt.json`
- `run-record.json`
- the saved output image

## Current nightly default

- family: `flux2klein`
- workflow: `models/flux2klein/image-edit-9b-distilled/workflow.json`
- output policy: single image only
- test style: smoke test, not broad sweep

## Review rule

Do not call a workflow fully optimal from one benchmark image. Use these batches to decide whether a path is promising enough for follow-up.

## Latest review

- 2026-08-27 overnight single-image smoke, reviewed: clean full-body studio restage with preserved blonde identity, gold corset, black thigh-high boots, and a clean white seamless background. Keep `runs/2026-08-27T02-08-31-134Z-single-smoke/` as the current benchmark winner for now, but queue one more smoke test before any broader promotion.
- 2026-08-23 overnight single-image smoke: clean full-body white-background restage with strong identity retention, a tidy silhouette, and no obvious background clutter. Keep `runs/2026-08-23T02-05-24-860Z-single-smoke/` as the current benchmark winner, but queue one more smoke test before any broader promotion.
- 2026-08-21 overnight single-image smoke: clean full-body studio restage on a pure white background, with good identity retention and a tidy silhouette. Treat it as the current benchmark winner for now, but keep one more smoke test queued before any stronger promotion.
- 2026-08-22 overnight single-image smoke: another clean white-background restage with strong identity retention, a tidy full-body silhouette, and no obvious background clutter. It was the prior winner before the 2026-08-23 run.