# FLUX image provenance

## Second series: photo stress tests, 2026-10-20

`boundary/` contains three fixed photo tasks run twice in the same six API models: 36 planned trials, 30 returned images and six trials without an image. One additional failed Max pose retry is also retained. No visual output was replaced or omitted. All returned originals were personally inspected before observations were written.

The full machine-readable provenance is [`boundary/test-record.json`](./boundary/test-record.json): frozen prompts and checks, source and processed-input hashes, exact submitted requests, model IDs, returned dimensions, reported seeds, output hashes, Gateway generation IDs, response cost receipts, failures and individual visual assessments. Source references are existing AI-created fictional adults and a pose sketch from this product. Display files are uncropped WebP quality 90; separately linked JPG/PNG files preserve returned bytes. Successful response charges total $1.592; failed calls without receipts are excluded, not assumed free. These are sample provider charges, not customer credit prices.

The task design draws on original external tests. Configuration differences and original sources are linked in the [photo-test methodology](https://photogenerai.com/blog/flux-image-models#photo-method). Our outputs are our own tests, not reproductions of an official benchmark. The original first-series provenance follows.

## First series: 2026-10-14

These images were generated for PhotoGenerAI on 2026-10-14 through its existing Vercel AI Gateway credential. They are demonstration assets, not customer submissions.

## Text prompt

A cobalt blue ceramic teapot with a curved handle, on a pale peach pedestal, soft studio lighting, cream background, product photography. No text.

## Generation receipts

| File                 | Model               | Gateway generation             | Actual USD cost | SHA-256                                                          |
| -------------------- | ------------------- | ------------------------------ | --------------- | ---------------------------------------------------------------- |
| flux-3-image.webp    | bfl/flux-3-image    | gen_01M4WNVNDYBX1RG3CDSNKDA5RW | 0.048           | f25fba743df2aac726fe97c9cc4731add360bc9f5b50f899c7e698a0c6a475a2 |
| flux-2-pro.webp      | bfl/flux-2-pro      | gen_01M4WNWGKFJ79TG4ZEGCK9KG3Z | 0.030           | 26c845fdc92f02815c82aebdd6e09db4dd7fc134b480d81e1a40294c4a11bfbf |
| flux-2-max.webp      | bfl/flux-2-max      | gen_01M4WNWST30NKYX8WCP3BM8N6C | 0.070           | 4c366bee1f97629166b8989c47717cf0bb1fab880b8a3f4d155a67347ab959a6 |
| flux-2-flex.webp     | bfl/flux-2-flex     | gen_01M4WNXDQ2B7YEB4HZ3D70554W | 0.050           | 9eb81330d8f06d2bb0b8eba30d2c81c54908f59998588cba79b0b7a1ad492615 |
| flux-2-klein-4b.webp | bfl/flux-2-klein-4b | gen_01M4WNXVAT4KCN9228Z0Y2HCMD | 0.014           | 8aaad70b30c025e496af9e2791e122479e6bb3d199640e3e249503ede3d51b0f |
| flux-2-klein-9b.webp | bfl/flux-2-klein-9b | gen_01M4WNY26V84NQXZ1QG9JC652Y | 0.015           | 4c82fa4bdceb5a6cabb5a6b7c133673475908573ce357d4e5a166265bdb6a0d7 |
| flux-2-pro-edit.webp | bfl/flux-2-pro      | gen_01M4WNY70FNF8XS74W4CH1JPK0 | 0.045           | 4edebe9abd1dec7c0d9ca7fe97b7dbeb7c42ea81564533d66d29c210f2711d3c |

## Photo edit

Input: `flux-2-pro.webp`. Prompt: “Change the ceramic teapot to deep forest green. Keep the peach pedestal and the cream background.” The input was resized to at most 960 pixels on its long edge, with a 1024 × 768 output.

Text generations use a square 1K output. FLUX 3 requests `resolution: "1k"`; FLUX.2 requests explicit width and height. All seven final requests returned without parameter warnings. Assets were encoded as WebP at quality 86.