tiger-ai-lab/rewardharness/artifact-and-structure-preservation

Artifact and Structure Preservation

A successful image edit must preserve the original image's structure and subject identity (unless instructed to change them) and avoid introducing severe artifacts.

• Over-editing: If an edit completely replaces the background or subject in a way that loses the original image's context, penalize it heavily.
• Artifacts vs. Instruction Following: If Image B follows the instruction (e.g., rotates an object, adds a crowd) but introduces severe unnatural distortions, missing parts, or obvious AI artifacts, while Image A fails the instruction but remains a clean, natural image, Image A is often the Ground Truth preference.
• Rule of Thumb: Do not reward an image for following the instruction if it destroys the visual quality, realism, or plausibility of the scene.

<!-- Release review: RewardHarness v0.4.1 (2026-08-26). -->