A critical analysis of semantic attribute leakage in compositional text-to-image generation
Text-to-image diffusion models suffer from a persistent and under-examined failure: semantic attribute leakage , wherein a visual attribute specified for one object—color, texture, or shape—erroneously manifests on a co-occurring object in the generated image. Distinct from catastrophic neglect and attribute swapping, leakage is a directional, cross-object contamination arising from three independent sources: entangled CLIP token representations, spatially unbound cross-attention maps, and training-data priors learned during pretraining. Despite its prevalence, leakage has never been subjected to a dedicated critical analysis as a standalone phenomenon with its own taxonomy, mechanistic framework, and benchmark critique. We address this gap by introducing a three-axis taxonomy (source, attribute type, severity) that unifies the fragmented terminology across the literature. We critically analyse the mechanistic roots of leakage, evaluate existing mitigation methods against the taxonomy, and assess whether current benchmarks actually isolate leakage from compositional failure broadly. Our analysis reveals that reported performance on texture and shape binding consistently lags color across every method with published scores on all three, that attention-manipulation methods offer limited gains at disproportionate cost, and that the transition to Diffusion Transformer architectures has shifted but not resolved leakage’s structural origin.
Authors
- Yogesh Kakde (ORCID: https://orcid.org/0000-0002-1943-3768)
- Sandeep Kumar Dash (ORCID: https://orcid.org/0000-0002-6237-1371)
- Ranjita Das (ORCID: https://orcid.org/0000-0001-6184-6294)
Institutions
- National Institute of Technology Agartala (IN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1038/s41598-026-79004-0
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- article
- Field-Weighted Citation Impact
- 0.00