Learning an Image Editing Model without Image Editing Pairs

TL;DR

This work introduces NP-Edit, a no-paired-data paradigm for training image-editing diffusion models by harnessing differentiable feedback from Vision–Language Models (VLMs) and enforcing realism through Distribution Matching Distillation (DMD). By unrolling a few diffusion steps and using VLM-based edit-verification and identity checks, the method learns $G_{ heta}$ without ground-truth edited images, achieving competitive results with supervised baselines and surpassing RL-based approaches when using the same VLM as reward. The combination of VLM guidance and DMD enables effective instruction-following across local and free-form edits while maintaining image realism, demonstrated on GEdit-Bench, ImgEdit, and DreamBooth customization tasks. The work highlights the potential of unpaired, VLM-guided fine-tuning for flexible, efficient image editing at scale, with clear ablations showing the necessity of both components and a path for future improvements via stronger VLMs and perceptual constraints.

Abstract

Recent image editing models have achieved impressive results while following natural language editing instructions, but they rely on supervised fine-tuning with large datasets of input-target pairs. This is a critical bottleneck, as such naturally occurring pairs are hard to curate at scale. Current workarounds use synthetic training pairs that leverage the zero-shot capabilities of existing models. However, this can propagate and magnify the artifacts of the pretrained model into the final trained model. In this work, we present a new training paradigm that eliminates the need for paired data entirely. Our approach directly optimizes a few-step diffusion model by unrolling it during training and leveraging feedback from vision-language models (VLMs). For each input and editing instruction, the VLM evaluates if an edit follows the instruction and preserves unchanged content, providing direct gradients for end-to-end optimization. To ensure visual fidelity, we incorporate distribution matching loss (DMD), which constrains generated images to remain within the image manifold learned by pretrained models. We evaluate our method on standard benchmarks and include an extensive ablation study. Without any paired data, our method performs on par with various image editing diffusion models trained on extensive supervised paired data, under the few-step setting. Given the same VLM as the reward model, we also outperform RL-based techniques like Flow-GRPO.

Figures (12)

• Figure 1: Method. We fine-tune a pretrained text-to-image model into a few-step image-editing model using differentiable VLM-feedback regarding edit success. In addition, we use distribution matching loss (DMD yin2024improved) to ensure output images remain in the natural image manifold.
• Figure 2: Qualitative comparison on GEdit-Bench under the few-step sampling setting. For an upper-bound comparison, in the $1^{\text{st}}$ column we show results of the best multi-step sampling method (as measured by the quantitative metrics in Table \ref{['tbl:metrics_editing']}). Our method performs on par or better than baseline methods across different edit types in the few-step setting. We show more samples in the Appendix Figure \ref{['fig:results_editing_gedit_supp']}
• Figure 3: Qualitative comparison on Customization task. Our method can generate the object in new contexts while having better fidelity under few-step sampling. We show more samples in the Appendix Figure \ref{['fig:results_editing_feeform_supp']}.
• Figure 4: Dataset and VLM scale and comparison with Reinforcement Learning on the GEdit-Bench. Increasing dataset scale and using stronger VLMs leads to increased performance. Our method also performs better than post-training an SFT model with RL liu2025flow.
• Figure 5: Performance for each sub edit-type. Training with only DMD loss fails to achieve certain tasks like removal and style changes. In addition, using binary cross-entropy loss and VLM identity-based questions helps improve the overall performance.