MultiBind: A Benchmark for Attribute Misbinding in Multi-Subject Generation
This addresses a key failure mode in multi-subject image generation for users needing fine-grained control, though it is incremental as it focuses on benchmarking rather than solving the problem directly.
The paper tackles the problem of cross-subject attribute misbinding in multi-subject image generation, where attributes are incorrectly assigned to wrong subjects, and introduces the MultiBind benchmark with a dimension-wise confusion evaluation protocol to diagnose such failures, revealing issues that conventional metrics miss.
Subject-driven image generation is increasingly expected to support fine-grained control over multiple entities within a single image. In multi-reference workflows, users may provide several subject images, a background reference, and long, entity-indexed prompts to control multiple people within one scene. In this setting, a key failure mode is cross-subject attribute misbinding: attributes are preserved, edited, or transferred to the wrong subject. Existing benchmarks and metrics largely emphasize holistic fidelity or per-subject self-similarity, making such failures hard to diagnose. We introduce MultiBind, a benchmark built from real multi-person photographs. Each instance provides slot-ordered subject crops with masks and bounding boxes, canonicalized subject references, an inpainted background reference, and a dense entity-indexed prompt derived from structured annotations. We also propose a dimension-wise confusion evaluation protocol that matches generated subjects to ground-truth slots and measures slot-to-slot similarity using specialists for face identity, appearance, pose, and expression. By subtracting the corresponding ground-truth similarity matrices, our method separates self-degradation from true cross-subject interference and exposes interpretable failure patterns such as drift, swap, dominance, and blending. Experiments on modern multi-reference generators show that MultiBind reveals binding failures that conventional reconstruction metrics miss.