Wei Xiong

We propose Multi-Modal Image Generation Benchmark (MMIG-Bench), a comprehensive benchmark for evaluating multi-modal image generation models. MMIG-Bench unifies compositional evaluation across T2I and customized generation, introduces explainable aspect-level metrics, and provides extensive human and automatic evaluations.

We propose an approach that improves the identity fidelity of generated objects by automatically locating and aligning visual tokens in a reference image with the target region to be refined.

We introduce GroundingBooth, a framework that achieves zero-shot instance-level spatial grounding for both foreground subjects and background objects in text-to-image customization.

We introduce SwapAnything, a framework that swaps arbitrary objects in an image with personalized concepts provided by reference images while preserving the surrounding context.

Our work achieves advanced image composition with strong identity preservation, automatic object viewpoint and pose adjustment, color and lighting harmonization, and shadow synthesis in a single framework.

We present PHOTOSWAP, an approach for personalized subject swapping in existing images while preserving the surrounding scene.

My Ph.D. thesis summarizes my research on guidance-driven visual content creation, including visually pleasing data synthesis and synthesis for downstream visual recognition tasks.

We introduce Image Sentiment Transfer, an important but still underexplored research task, and propose an effective and flexible framework that performs sentiment transfer at both the image and object levels.

We transform images from fine-grained categories to synthesize new examples that preserve the identity of the input, thereby benefiting fine-grained recognition and few-shot learning.