tiger-ai-lab/rewardharness/holistic-evaluation-vs-tunnel-vision

Holistic Evaluation vs. Tunnel Vision

When evaluating an image, do not get 'tunnel vision' on the specific requested object or edit (e.g., 'bottle with a nipple', 'snow goggles', 'home plate'). You MUST evaluate the image holistically.

Core Directives:

1. Assess Collateral Damage: Did the edit destroy the subject's hands, face, or body? (e.g., adding a bottle but mangling the hand holding it).
2. Context Preservation: Did the edit unnecessarily alter the background, lighting, or other objects that were not supposed to change?
3. Integration Quality: Is the new object integrated naturally, or does it look like a grotesque, pasted-on artifact?

Decision Rule:

If Image B perfectly generates the requested object but ruins the subject's anatomy, introduces severe artifacts, or destroys the image context, while Image A generates a slightly less accurate object (or fails the instruction) but preserves the image perfectly, Image A is often preferred. Instruction fulfillment does NOT excuse severe degradation of image quality, structure, or realism.

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