CausalSliders: Graph-Guided LoRA Interventions for Causally Consistent Image Editing

Text-to-image diffusion models enable flexible semantic editing but lack causal control: the ability to intervene on a target factor while preserving causally independent attributes during editing. Trained observationally, these models encode both causal and incidental correlations. For example, increasing age should affect wrinkles, a causal descendant, but not camera pose or background, which may be correlated yet causally independent. Existing editing methods entangle these effects because optimizing target accuracy under correlated factors exploits spurious co-occurrences, trading invariance for fidelity and failing to enforce causal mediation. We introduce CausalSliders, a parameter-efficient image-editing framework that embeds causal structure directly into diffusion-model adaptation. We represent each semantic factor by a dedicated low-rank LoRA adapter trained to induce a targeted parameter-space effect, enabling reusable edits without per-image optimization. Minimality and conditional-independence losses penalize non-target drift and cross-factor interference, addressing failures of linear, commutative multi-LoRA composition under correlated factors. Factor dependencies are encoded by a directed acyclic graph, and a gated, non-commutative composition operator applies interventions in causal order to enforce mediation under multi-attribute edits. By enforcing graph-ordered parameter interventions, CausalSliders improves multi-factor correctness from 39% (Concept Sliders) to 72% and achieves 81% causal path accuracy, matching the causal accuracy of Deep-SCM while running over 50× faster without per-image optimization.