A prompt that works in one AI image generator may fail in another because interfaces apply different defaults, hidden prompt enhancement, safety rules, aspect-ratio handling, and model routing. Copying a prompt between tools is therefore not a neutral comparison.

This five-prompt test measures portability without rewriting the prompt for each route.

Freeze the test conditions

Before generating, record the exact prompt, negative prompt, aspect ratio, reference images, output count, seed controls, and visible quality settings. Disable automatic prompt enhancement when the interface allows it. If it cannot be disabled, record that limitation instead of pretending the routes are identical.

Generate four outputs per prompt. Keep the first result from every run and do not replace weak outputs with hand-picked winners.

Prompt 1: composition

Request a subject in a precise area of the frame and reserve empty space elsewhere. Example requirements might place a ceramic bottle in the lower-right third while keeping the upper-left quadrant clear for copy.

Score subject position, subject scale, and whether the reserved area remains genuinely usable.

Prompt 2: object relationships

Describe three objects with explicit relative positions, such as a red cube behind a glass sphere and a small brass key to the sphere's left.

Score object count, color, depth order, and left-right relationships. This prompt exposes models that preserve nouns but lose spatial instructions.

Prompt 3: material and lighting

Specify a material whose appearance depends on highlights and texture, then define one light source. Brushed aluminum under a large softbox is a useful example because the material should show directional grain without becoming chrome.

Score material identity, highlight shape, shadow direction, and whether the lighting remains coherent across the object.

Prompt 4: readable text

Ask for one short word on a sign, package, or poster. Keep the word between four and seven letters and avoid stylized spelling.

Score exact characters, character order, placement, and whether the surrounding design stays intact. Record near misses as failures when exact text is mandatory.

Prompt 5: a real production brief

Use one recent brief from the team's backlog. Remove customer names and confidential details, but preserve the actual composition, brand constraints, delivery format, and intended use.

This prompt prevents the benchmark from optimizing only for artificial test cases.

Use one scoring sheet

For each output, score the following from zero to two:

  • instruction compliance;
  • composition stability;
  • object and identity accuracy;
  • artifact severity;
  • production readiness.

Zero means unusable, one means recognizable but requiring material correction, and two means acceptable without a production-blocking defect. The maximum is ten points per image.

Also record the first failure category: omitted instruction, spatial error, identity drift, text error, malformed geometry, unwanted artifact, export limitation, refusal, or timeout. A total score without failure labels hides why a route succeeds or fails.

Compare without prompt tuning

Run the unchanged five-prompt set through every candidate route. Teams can include a free AI image generator as one disclosed test environment while keeping prompts, settings, output counts, and scoring rules fixed. The link is not evidence that the generator will outperform another route.

Report the median score, acceptance rate, and dominant failure category for each prompt type. Do not declare a universal winner. A route that is strong on materials may still be weak on typography or spatial relationships.

Add a second round only after the baseline

After the frozen baseline, allow one route-specific prompt revision. Record every changed word and compare the improvement against the added tuning time. This separates raw portability from the skill required to adapt a prompt to a particular interface.

Date every run and retain the displayed model label. Retest after routing, default enhancement, model, or export settings change.

Disclosure: Written and published by PhotoArtify Team on September 14, 2026. We operate PhotoArtify and may benefit if readers use the linked image generator. This protocol is a reproducible evaluation method, not an independent endorsement or a universal quality claim.