CheckDistill improves reward models for guided image editing
CheckDistill: Adaptive Checklist Distillation for Efficient Distributional Reward Modeling in Image Editing
Computer Vision and Pattern Recognition
Summary
When computers edit images by following instructions, they need feedback that tells them how well they did. The authors created CheckDistill, a method that breaks down editing goals into smaller checks to provide clearer and more efficient scores. This system learns by watching a teacher model explain why certain edits are good or not. Once trained, CheckDistill quickly rates how well an image was edited without needing the extra explanations, making the feedback both accurate and fast.
What this means in practice
- •For ai developers for image editing: Enable efficient and accurate reward feedback to improve reinforcement learning models that guide image editing with user instructions.
- •For computer vision engineers: Deploy fast, reliable evaluation models that score image edits without costly reasoning, helping optimize image manipulation pipelines.
Authors
Yaqi Cai, Zuan Gao, Mengting Chen, Wei Li, Bo Zheng
Abstract
Reinforcement learning for instruction-guided image editing requires reward models that provide reliable, accurate, and efficient feedback. Direct scalar prediction compresses multidimensional editing preferences, whereas generative evaluators can expose their judgment through explicit reasoning. However, broad scoring criteria do not specify all instance-dependent editing and preservation constraints, and generating long reasoning trajectories is costly for online optimization. We introduce CheckDistill, a pointwise reward modeling framework that internalizes instruction-adaptive checklists. We decompose each editing condition into atomic, independently verifiable requirements and construct four-dimensional scoring labels together with structured verification evidence. After supervised score initialization, a privileged teacher conditions on checklists, item-level judgments, and dimension-level rationales to supervise the student's own scoring trajectories. A reliability-weighted on-policy self-distillation objective adjusts supervision at score tokens according to agreement between the teacher's expected scores and reference labels, while retaining candidate probability weights in a truncated token support. The deployed student directly predicts four score distributions without checklist inputs or explicit reasoning and aggregates their expectations into a scalar reward. CheckDistill achieves aggregate accuracies of 72.30%, 56.67% and 82.90% on MMRB2, MER-Bench and EditReward-Bench, respectively, exceeding the compared open-source baselines on all three metrics.