lijiayangCS/StableI2I_PLUS

StableI2I-PLUS: Image-to-Image Transition Evaluation

StableI2I-PLUS is an 8 billion parameter model developed by lijiayangCS, designed as a unified and dynamic evaluation framework for image-to-image (I2I) transitions. Accepted by ICML 2026, its core purpose is to measure content fidelity and pre-post consistency, addressing the common limitation of existing I2I evaluations that often overlook how faithfully an output image preserves the input's semantic correspondence, spatial structure, and low-level appearance.

Key Capabilities

• Content Fidelity Assessment: Evaluates unintended changes across three levels without requiring reference images:
  • Semantic Level: Detects object addition, removal, replacement, or identity drift.
  • Structure Level: Identifies misalignment, deformation, repainting, and structural distortion.
  • Low-level Appearance: Checks for visual degradations like blur, noise, color cast, or artifacts.
• Explicit Fidelity Score: This specific checkpoint provides a direct numerical score summarizing the overall content consistency between input and output images.
• Broad Applicability: Can be applied to various I2I tasks, including image editing and image restoration.

Good For

• Developers needing a quantitative measure of content consistency in I2I outputs.
• Evaluating the preservation of semantic, structural, and low-level details in generated images.
• Use cases where a direct numerical fidelity score is preferred over fine-grained diagnostic outputs (though a slight degradation in fine-grained accuracy compared to the paper's original checkpoint is noted).

For the latest code, demos, and inference scripts, refer to the official GitHub repository.