BokehDiff: Neural Lens Blur with One-Step Diffusion

2025-07-23

AI Summary

Existing lens blur rendering methods suffer from depth estimation errors, leading to artifacts at depth discontinuities and struggling to balance physical accuracy with visual plausibility. To address this, we propose a generative diffusion-prior-based neural lens blur rendering framework. Our method introduces a physics-inspired, depth-aware self-attention mechanism that explicitly models circle-of-confusion scaling and self-occlusion. We adopt a single-step diffusion inference paradigm, enabling efficient, high-fidelity blur synthesis without iterative denoising. Furthermore, we leverage diffusion models to synthesize diverse foreground images with alpha mattes, constructing physically aligned training data. Experiments demonstrate that our approach significantly outperforms state-of-the-art methods on both synthetic and real-world scenes. Notably, it achieves sharper, more natural blur transitions at object boundaries and depth edges—delivering superior fidelity while maintaining real-time inference potential.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Generation
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Search and Retrieval-Augmented AI: Personalized, context-aware and across-device search
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Security and Privacy: Data transparency and provenance

Abstract

We introduce BokehDiff, a novel lens blur rendering method that achieves physically accurate and visually appealing outcomes, with the help of generative diffusion prior. Previous methods are bounded by the accuracy of depth estimation, generating artifacts in depth discontinuities. Our method employs a physics-inspired self-attention module that aligns with the image formation process, incorporating depth-dependent circle of confusion constraint and self-occlusion effects. We adapt the diffusion model to the one-step inference scheme without introducing additional noise, and achieve results of high quality and fidelity. To address the lack of scalable paired data, we propose to synthesize photorealistic foregrounds with transparency with diffusion models, balancing authenticity and scene diversity.

Problem

• Achieves accurate lens blur using diffusion prior
• Reduces artifacts in depth discontinuities via physics-inspired module
• Synthesizes photorealistic data to address paired data scarcity

Innovation

• Physics-inspired self-attention for image alignment
• One-step diffusion without additional noise
• Synthetic photorealistic foregrounds with diffusion models