Stack multiple AI character adapters onto one diffusion model and you often get a blurry mess of mixed-up faces and traits instead of a clean scene. A new method called SDO, short for Subspace Deconflicting Operator, tries to fix that at the adapter level rather than by retraining anything.
The paper, posted to arXiv on August 17, 2026, describes a problem familiar to anyone composing scenes from independently trained adapters: combine several of them in one generation and characters start borrowing each other's faces, clothes, or features, while the overall scene composition gets shakier. The authors trace this to overlapping "dominant subspaces" inside the adapters' shared layers, essentially competing internal directions that clash when adapters are stacked. SDO reconstructs each adapter's low-rank updates, extracts a compact signature of its subspace, and measures how much those signatures overlap with each other. It then applies a transformation that suppresses the conflicting shared directions while preserving what makes each adapter distinct, before converting everything back into standard adapter updates that plug into existing diffusion inference pipelines.
Multi-adapter composition is the practical way most tools generate scenes with several distinct characters, since training one custom model per character combination doesn't scale. Right now, creators mostly work around interference by limiting how many adapters they stack or hand-tuning results. The paper reports that SDO's improvements to identity fidelity and compositional stability get more pronounced as more adapters are composed together, which is where current pipelines tend to struggle most.
As with most single-paper fixes, the real test is whether SDO holds up on adapters and characters the authors didn't benchmark, and whether it's worth the added inference-time computation.