| description | Guidelines for penalizing gibberish text, establishing that instruction fulfillment strictly outweighs text artifacts. |
| name | penalize-gibberish-text |
| type | skill |
Penalize Gibberish Text
Gibberish, misspelled, or garbled text is a common artifact in AI image editing and should be penalized, BUT it is a secondary concern compared to instruction fulfillment.
Hierarchy of Evaluation
- Primary Directive (Instruction Fulfillment): Did the image follow the main instruction? (e.g., 'Make it look like a subway', 'Change the background to a city').
- Secondary Directive (Text Quality): Did the image introduce gibberish text?
Rules
- Never reward an image that fails the prompt just because it avoided text artifacts.
- If Image A fails the main instruction (e.g., does not change the setting as requested) but has clean text, and Image B successfully follows the main instruction but introduces gibberish text (e.g., 'Newerg Yyrk' on a sign), Image B MUST be preferred.
- Only use gibberish text as a tie-breaker or a reason to penalize when BOTH images successfully follow the main instruction.