Image models generating partially-eaten burritos over time (www.generativist.com)

Researchers and hobbyists ran a simple but revealing probe: feed a state‑of‑the‑art image generator (fal.ai defaults) a single prompt — “A partially eaten burrito with cheese, sour cream, guacamole, lettuce, salsa, pinto beans, and chicken” — and inspect many outputs. Rather than reliably reproducing a believable cross‑section, the model frequently produced smushed, congealed interiors and inconsistent ingredient placement; repeated generations varied widely. The exercise, inspired by meme‑era visual probes and benchmarks like Simon’s Pelican work, highlights how even common, well‑represented objects in training data can be hard to render when they involve occlusion, non‑rigid deformation and messy internal structure. This is significant because it exposes concrete failure modes of modern generative vision models: poor compositionality, sensitivity to occlusion and texture blending, and dataset gaps for partially‑destroyed or highly variable instances. Technically, the experiment used default model settings (eschewing heavy prompt engineering/HIL), so variability reflects the model’s learned priors rather than crafty prompting. The result suggests straightforward follow‑ups for the AI/ML community: curate more examples of “broken” or occluded objects, develop benchmarks that test non‑rigid composition and internal structure, and measure robustness across repeated samplings — important for applications in food imagery, forensics, and any use case demanding coherent internal object representations.