PreserveChange: Disentangled Preservation and Change Control for Instruction-Guided Image Editing

Instruction-guided image editing with diffusion models allows intuitive, language-based modifications but often suffers from inconsistent results, edits may distort backgrounds or fail to apply sufficiently. We propose DO-ELECT (Dual-Objective ELECT), an extension of the ELECT framework that explicitly balances two goals: preserving non-target regions and ensuring adequate target-region edits. DO-ELECT introduces a Dual Background Inconsistency Score (Dual-BIS) to jointly penalise unintended background changes and reward effective target modifications during early-timestep seed selection. A strict preservation mode further guarantees pixel-perfect background consistency via learned mask compositing. Evaluated on 100-image subsets from MagicBrush and PIE-Bench using InstructPix2Pix and MagicBrush backbones, DO-ELECT consistently improves preservation metrics (SSIM, MSE) over baseline and single-objective ELECT while maintaining comparable instruction-following performance (CLIP score).

• ›Dual-BIS: a seed selection metric that jointly optimises for preservation and change objectives, addressing the under-editing limitation of single-objective ELECT.
• ›Strict preservation mode guaranteeing pixel-perfect background consistency via learned mask compositing.
• ›Up to 17.5-point SSIM gains over baseline inference across MagicBrush and PIE-Bench.
• ›Consistent preservation improvements on two backbones (InstructPix2Pix, MagicBrush) at a CLIP-score cost under 0.008.

DO-ELECT extends ELECT by evaluating multiple diffusion seeds at early timesteps using Dual-BIS, a composite score that penalises background change and rewards target-region modification simultaneously. Spatial masks derived from the editing instruction separate background from target pixels. The strict preservation variant post-processes outputs by compositing edited target regions onto the original background using learned masks, guaranteeing zero background distortion.

100-image subsets, images resized to 512×512 before evaluation. SSIM and MSE measure background preservation; CLIP Score measures instruction following.

• ›Strict mode gives the strongest preservation everywhere, because it restores original pixels outside the inferred change region by compositing — SSIM 0.6639 → 0.8387 on MagicBrush with InstructPix2Pix.
• ›Dual-objective seed selection alone still improves on single-objective ELECT consistently, if modestly: SSIM 0.7023 → 0.7128, MSE 0.0169 → 0.0158.
• ›CLIP scores stay broadly comparable across methods; strict mode trades a little global image–text similarity for precise background preservation.
• ›Ablations: stopping step t=40 of 100 balances estimate quality against cost, 10 candidate seeds captures most of the benefit of 20, and the default weights (5.0, −2.0, 0.5) balance the two objectives.
• ›Runtime on an A100: 8s baseline, 12s for ELECT/DO-ELECT, 13s with strict preservation.