Open-domain image editing with FLUX and structural causal modeling

Open-domain image editing requires accurate semantic modification while preserving identity, structure, and non-target visual content. Existing diffusion- and rectified-flow-based editors often remain affected by attribute coupling, spurious correlations, and unintended changes to semantically related regions. This study presents a causal-structure-guided image editing framework built on FLUX rectified flow. A learnable hierarchical structural causal model (SCM) is introduced to represent directional dependencies among latent semantic attributes through an adaptive adjacency matrix constrained by acyclicity and sparsity. The resulting structural representation is integrated into the Multi-Modal Diffusion Transformer (MM-DiT) dual-stream attention through attribute-to-token projection. A counterfactual consistency objective based on the abduction–intervention–prediction formulation is further incorporated to align the intended semantic change with preservation of non-target information. Experiments on PIE-Bench, MagicBrush, CelebA-HQ, and a 20,000-sample LAION-derived editing subset demonstrate consistent improvements in editing fidelity and attribute specificity. On PIE-Bench, the proposed method improves Contrastive Language-Image Pre-training (CLIP) text alignment (CLIP-T) by approximately 1.7% and reduces Learned Perceptual Image Patch Similarity (LPIPS) by 5.1% relative to FluxSpace. On CelebA-HQ, compared with FluxSpace, the Attribute Cross-Interference Rate (ACIR) decreases from 0.2074 to 0.1763, while identity similarity increases from 0.7692 to 0.7854. Ablation experiments further verify the complementary contributions of structural modeling, causal attention intervention, and counterfactual consistency. In the query, key, and value masking ablation study, full masking increases inference time by approximately 3.7% relative to the unmasked configuration. The framework provides a practical approach to controllable and identity-preserving open-domain image editing.