SDO Resolves Conflicts in Multi-Adapter Diffusion Models.

When generating images with multiple characters or complex scenes using diffusion models, a common approach is to compose several independently trained adapters within a shared backbone. However, this often leads to undesirable outcomes such as identity mixing, where character features blend, cross-character attribute leakage, and overall unstable scene compositions. Researchers hypothesize that this interference stems from conflicts between overlapping dominant subspaces within the shared layers of the diffusion model. To mitigate this, they propose the Subspace Deconflicting Operator (SDO). SDO works by reconstructing layer-wise low-rank updates from the selected adapters and extracting compact subspace signatures. The operator then measures pairwise conflicts by analyzing the overlap in output subspaces and applies a permutation-equivariant transformation. This transformation effectively suppresses harmful shared directions while carefully preserving the unique, identity-specific characteristics of each adapter. The modified representations are then re-mapped into standard adapter updates, allowing direct integration into existing diffusion inference pipelines. Experimental results consistently show that SDO significantly enhances identity fidelity and compositional stability, with particularly noticeable improvements as the number of simultaneously composed adapters increases.