tiger-ai-lab/rewardharness/environmental-blending-and-lighting-adaptation

Environmental Blending and Lighting Adaptation

When an instruction requests a drastic change in the background or environment (e.g., moving polar bears from ice to a grassy savannah, or changing a scene from day to night), the subject must be integrated into the new environment.

Guidelines

1. Expect and Reward Adaptation: The subject should undergo changes in lighting, color temperature, and shadows to match the new environment. For example, a subject moved to a sunny savannah should have warmer lighting; a subject moved to a dark forest should have appropriate shadows.
2. Ground Contact and Physics: The way the subject interacts with the ground must change. Paws on flat ice look different from paws standing in tall grass. Do not penalize an image for "altering the subject" if the alteration is simply grass covering the feet, or a change in shadow casting.
3. Integration over Strict Pixel Preservation: While the core identity and structure of the subject must be preserved, strict pixel-perfect preservation of the subject from the original image is often a negative if it results in a "cut-and-paste" look. An image that successfully blends the subject into the new environment (even if it slightly alters their lighting or edges) is preferred over one that perfectly preserves the original pixels but fails to integrate them, leaving them looking unnatural in the new setting.

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