Style Recipe Tests

Style Recipes are ChatGPT instructions that ensures image generation outputs in a specific style.

Each Style Recipe is rigorously tested in the scenarios described below.

Test conditions

1 prompt attempt to reach the desired result
No resets between prompts are allowed, unless test score falls below 85%
Increasing difficulty and detail in each prompt → From trivial to very hard
Expect characters to maintain consistency, and document when drift occurs
Each test is scored from 0-100%. Examples:
- 100% = perfectly consistent match
- 95% = minor missing details
- 90% = some missing details, drift or hallucination, but character could otherwise pass as the same initial base character.
- 85% = major details missing, major drift or hallucination. Character otherwise consistent, but requires a reset.
- 80% = character visibly inconsistent with the base character. Reset required.

<aside>

Scores are provided to help you scan performance. All results are included, scores are subjective assessments — you can judge for yourself.

For the last 3 tests of each test suite, a new conversation was started, due to the hard limit of 20 uploads per conversation with ChatGPT, to ensure fairness across all test suites (I’ve later found out the hard limit is 100 uploads).

</aside>

Whenever a reset is mentioned, I used the reset method described in the ‘ChatGPT Style Consistency Toolkit’ — you can do that too!

Prompt template tested:

[Upload anchor image]

Action: [Describe what the character is doing]

Pose: [Describe body language]

Expression: [Emoji or mood]

Props: [Optional objects interacting with the character]

Outfit: [Optional changes to the characters outfit]

Scene: [Describe location]

Additional notes: [Background, lighting, styling]

Curious about the workflow I used to get these results?