Negative-Guided Subject Fidelity Optimization for Zero-Shot Subject-Driven Generation

We present Subject Fidelity Optimization (SFO), a novel comparative learningframework for zero-shot subject-driven generation that enhances subjectfidelity. Beyond supervised fine-tuning methods that rely only on positivetargets and use the diffusion loss as in the pre-training stage, SFO introducessynthetic negative targets and explicitly guides the model to favor positivesover negatives through pairwise comparison. For negative targets, we proposeCondition-Degradation Negative Sampling (CDNS), which automatically generatesdistinctive and informative negatives by intentionally degrading visual andtextual cues without expensive human annotations. Moreover, we reweight thediffusion timesteps to focus finetuning on intermediate steps where subjectdetails emerge. Extensive experiments demonstrate that SFO with CDNSsignificantly outperforms baselines in terms of both subject fidelity and textalignment on a subject-driven generation benchmark. Project page:https://subjectfidelityoptimization.github.io/