03 / MODELS, RESULTS & CONFIDENCE
Models and architectures
Nine vision LLMs and five specialized counting or detection models, evaluated on the same six images. The LLMs are also deep-learning models; “specialist” describes their task, not a separate branch of AI.
Checking the saved results…
LLM vision backbones and parameter counts are not recorded in these evaluation files, so no specific CNN or Transformer backbone is assigned to them. DL architectures come from their saved configurations and official implementations.
Conclusions and interpretation
These are observations from six selected images, not a general model ranking. Three images are in FSC-147’s training split for dataset-trained checkpoints. FamNet receives visual examples; the other saved DL runs use text only. Neither comparison isolates architecture as the cause.
What stated confidence actually meant
Nine evaluated LLMs only · each point is an occupied confidence band · larger dots mean more answers
Hover, tap or keyboard-focus a dot for its model and sample count. Use the selector to separate overlapping points.
Which models counted more closely?
Which LLM confidence scores matched the outcomes?
All six images per LLM · sorted by observed Brier score, lower first
Mean stated confidence Exact-match accuracy
Methods, configurations and source files
LLMs: the original 3 October 2026 run only, joined to supplied Group 1 counts by group and image. Local models: the five current successful CPU runs, with configurations and dependency/checkpoint hashes in their JSON. No new inference requests are made by this page.
Confidence means P(exact count), not probability that an estimate is close. DL object scores are not comparable to that probability, so DL models are excluded from both confidence charts. CPU latencies and hosted LLM request latencies are different measurement settings; no speed or cost winner is inferred.
Dataset: FSC-147 / Learning To Count Everything, Viresh Ranjan, Udbhav Sharma, Thu Nguyen and Minh Hoai (CVPR 2021). Source images · Paper · Recorded download URL. Supplied counts have not been independently audited here.
DL setup and verification · Dataset split provenance · Scoring metadata · LLM runner
Presentation guidance: data-story and data-viz. The confidence chart follows the reference method, using this project’s saved answers.