<< All versions

Skill v1.0.0

currentAutomated scan100/100
tiger-ai-lab/rewardharness/artifact-and-structure-preservation
──Details
PublishedNovember 3, 2026 at 07:29 AM
Content Hashsha256:bbcc920eba3e1446...
Git SHA0c1e7051ab4f
──Files
Files (1 file, 1.1 KB)
SKILL.md1.1 KBactive
SKILL.md · 19 lines · 1.1 KB

schema_version: 2 name: artifact-and-structure-preservation kind: skill description: Rules for prioritizing image integrity and penalizing severe artifacts or loss of original structure. metadata: version: "1.0.0"


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). -->

All versions