Compare Image Models Without Fooling Yourself
One impressive example proves very little. A useful comparison controls the prompt, repeats the test, and scores what matters for the actual job.
Hold the Brief Still
- Choose a real task and write observable pass criteria before generating anything.
- Use the same source image, prompt, aspect ratio, and closest available quality setting.
- Run each model at least three times. Separate typical performance from a lucky result.
- Judge outputs without model labels when possible, then reveal the models after scoring.
- Record speed, cost, editability, policy constraints, and workflow friction alongside image quality.
Instruction Fit
Did it include the requested subject, action, composition, and exclusions?
Consistency
Did identity, product shape, typography, or reference details remain stable?
Craft
Are anatomy, perspective, light, edges, materials, and spatial relationships believable?
Usability
Can the result be revised, reproduced, delivered at the right size, and used under the applicable terms?
Task: [real use case]
Models: [A, B, C]
Runs per model: [number]
Criteria and weights: instruction fit [ ], consistency [ ], craft [ ], usability [ ]
Automatic failures: [list]
Winner: [model and evidence]
Uncertainty: [what this test did not establish]
Original Source
Inspired by Greptile's secondary overview of an Awesome Nano Banana repository and its emphasis on categorized, side-by-side model examples. This testing protocol is original and does not treat the overview as the owner of the underlying community content.