# GPT-Images-2 Feasibility Probe Notes

Date: 2026-06-10

## Method Interpreted From Project Docs

The project frames Vision Banana's idea as:

```text
input image + text/schema prompt -> generated RGB structure map -> decoded mask
```

For this probe, the target palette was:

- background: `#000000`
- target: `#FFFF00`

## Outputs

- `baby_person_rgb_mask_gpt_images_2.png`
  - Target: full visible baby/person.
  - Result: qualitatively successful as a natural-image foreground mask.
  - Caveat: output resolution differs from the input, and the mask is not guaranteed to be pixel-exact without post-hoc palette projection.

- `spine_rgb_mask_gpt_images_2_attempt1.png`
  - Target: visible vertebral column in the AP X-ray.
  - Result: failed as a medical segmentation output. The model generated an idealized spine-like structure on black background rather than a mask registered to the source X-ray.

- `spine_overlay_gpt_images_2_attempt2.png`
  - Target: visible vertebral column in the AP X-ray.
  - Result: better image anchoring because the X-ray is preserved, but still not a reliable segmentation. The overlay contains side-view vertebra-like shapes and does not strictly follow AP X-ray boundaries.

## Preliminary Feasibility Judgment

GPT-Images-2 can demonstrate the broad Vision Banana-style interface on ordinary natural images, but this zero-shot probe is not sufficient for medical imaging segmentation. For spine X-rays, the main failure is not color choice; it is coordinate fidelity and anatomy-grounded decoding. The model tends to generate plausible anatomy instead of extracting a strictly registered mask from the image.

For academic follow-up, GPT-Images-2 should be treated as a qualitative probe or upper-bound inspiration, not as the core reproducible method. A serious study should use ground-truth medical masks, strict RGB codec projection, Dice/HD95/surface metrics, negative prompts, uncertainty sampling, and baselines such as nnU-Net, MedSAM, SAT, VISTA3D, or SegVol.