ViT-512x382.jpeg -- a single-picture stress test for vision transformers
and vision-language pipelines. A 512x382 progressive JPEG, deliberately
dense in objects, textures, and text, with features aimed at the
structural mechanics of ViT-style models:

- Patch grid: a faint grid is superimposed over the whole scene,
  mimicking ViT patch extraction and probing spatial grid alignment.
- High-density signage: a large architectural sign structure carries an
  array of dozens of prohibition icons (including a dog-on-a-skateboard
  glyph), a "NO CATS ALLOWED" sign, and a "BEACH RULES & REGULATIONS"
  text panel -- dense visual and semantic tokens in one region.
- Serialized text: an airplane tows a banner of pseudo-token text
  (patch/grid/serialization themed), a token stream embedded in pixels.
- Semantic conflict: the no-cats rule and prohibition grid contradict
  the scene itself -- a cat rests on a car while anthropomorphic dogs
  surf, dig, and play across the beach, challenging detection and
  scene-understanding alike.
- Texture gradients: pavement, sand, ocean water, and foliage span the
  frame to probe spatial interpolation.
- Object inventory (one of each, minimum): cat, dog, person, airplane,
  bicycle, motorcycle, car, palm tree, beach, ocean, sun.

The scene reads as a parking lot in the foreground looking toward a
crowded beach and the ocean: cars, a motorcycle, and a bicycle in the
lot; a person with a dog; a thatched "DOGGY PADDLE RENTAL" hut with
surfboards; palms on the right; the sun and the banner plane in the sky.

A good single-image benchmark: OCR the signage, count or name the
objects, resolve the rule-vs-scene contradictions, or describe the
scene end to end.
