![[2602.06355] Di3PO - Diptych Diffusion DPO for Targeted Improvements in Image Generation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.06355] Di3PO - Diptych Diffusion DPO for Targeted Improvements in Image Generation
Summary
The paper presents Di3PO, a novel method for improving image generation in text-to-image diffusion models by efficiently creating targeted positive and negative image pairs.
Why It Matters
Di3PO addresses significant inefficiencies in current preference tuning methods for image generation, which can hinder the development of more accurate and effective AI models. By isolating specific areas for improvement, this approach enhances training efficiency and model performance, making it relevant for researchers and practitioners in the field of computer vision and AI.
Key Takeaways
- Di3PO improves the efficiency of preference tuning in image generation.
- The method creates targeted positive and negative image pairs, enhancing training quality.
- It showcases significant improvements in text rendering tasks over existing methods.
- The approach reduces computational costs associated with generating training pairs.
- Di3PO's design maintains surrounding context stability in images.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.06355 (cs) [Submitted on 6 Feb 2026 (v1), last revised 19 Feb 2026 (this version, v2)] Title:Di3PO - Diptych Diffusion DPO for Targeted Improvements in Image Generation Authors:Sanjana Reddy (1), Ishaan Malhi (2), Sally Ma (2), Praneet Dutta (2) ((1) Google, (2) Google DeepMind) View a PDF of the paper titled Di3PO - Diptych Diffusion DPO for Targeted Improvements in Image Generation, by Sanjana Reddy (1) and 4 other authors View PDF HTML (experimental) Abstract:Existing methods for preference tuning of text-to-image (T2I) diffusion models often rely on computationally expensive generation steps to create positive and negative pairs of images. These approaches frequently yield training pairs that either lack meaningful differences, are expensive to sample and filter, or exhibit significant variance in irrelevant pixel regions, thereby degrading training efficiency. To address these limitations, we introduce "Di3PO", a novel method for constructing positive and negative pairs that isolates specific regions targeted for improvement during preference tuning, while keeping the surrounding context in the image stable. We demonstrate the efficacy of our approach by applying it to the challenging task of text rendering in diffusion models, showcasing improvements over baseline methods of SFT and DPO. Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) Cite as: arXiv:2602.0...