lijiayangCS/StableI2I
StableI2I by lijiayangCS is an 8 billion parameter model designed as a unified and dynamic evaluation framework for image-to-image (I2I) transitions. It assesses content fidelity and pre-post consistency by evaluating semantic, structural, and low-level appearance changes without requiring reference images. This model is particularly useful for diagnosing unintended changes and benchmarking real-world I2I systems across various tasks like image editing and restoration.
Loading preview...
StableI2I: Evaluating Image-to-Image Transitions
StableI2I, developed by lijiayangCS, is an 8 billion parameter model accepted at ICML 2026, providing a novel evaluation framework for image-to-image (I2I) tasks. Unlike traditional methods that focus on instruction following or perceptual quality, StableI2I specifically addresses the critical issue of unintended changes in generated images, ensuring content fidelity and consistency without needing reference images.
Key Capabilities
- Comprehensive Evaluation: Assesses I2I transitions from three complementary perspectives:
- Semantic Level: Detects unintended object-level or meaning-level changes (e.g., object addition, removal, identity drift).
- Structure Level: Verifies preservation of spatial layout and geometric consistency (e.g., misalignment, deformation).
- Low-level Appearance: Identifies visual degradations like blur, noise, or color cast.
- Dynamic Framework: Applicable to a wide range of I2I tasks, including image editing and restoration.
- Benchmark Integration: Used to construct StableI2I-Bench, a benchmark for systematically evaluating MLLMs' ability to judge content fidelity.
- Human Correlation: Extensive experiments demonstrate strong correlation with human subjective judgments, providing accurate and interpretable evaluations.
When to Use This Model
StableI2I is ideal for developers and researchers who need to:
- Diagnose Content Consistency: Identify and understand unintended alterations in I2I outputs.
- Benchmark I2I Systems: Systematically evaluate and compare the fidelity of different I2I models.
- Ensure Quality Control: Verify that I2I models preserve semantic correspondence, spatial structure, and low-level appearance.
This HuggingFace repository hosts the checkpoint from the paper. For the latest code, inference scripts, and score-supported versions, refer to the official GitHub repository.