How this AI Learned to Refocus Any Photo

2025-12-20 · Source: Jia-Bin Huang · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Intermediate, long

Summary

AI Refocusing enables post-capture manipulation of photographic focus, allowing users to add or remove defocus blur, or shift focus to different subjects within a single image. The underlying technology builds upon traditional camera optics, explaining concepts like pinhole cameras, lenses, Gaussian lens formula, and the circle of confusion. The system trains two AI models: one to predict an all-in-focus image from a single-size defocused input, and another to generate desired defocus blur from an all-in-focus image and a target defocus map. This second model is fine-tuned using a pre-trained text-to-image model and real-world defocused images to ensure realistic blur effects. Applications include adjusting aperture size, changing the focus plane in group photos, sharpening small subjects, and even altering aperture shape for creative bokeh effects, offering significant creative freedom compared to traditional photography.

Key takeaway

For AI Engineers developing image manipulation tools, this refocusing model offers a robust approach to post-capture focus adjustment. You should consider integrating depth map estimation and a two-model architecture, leveraging pre-trained generative models for realistic blur synthesis. This can significantly enhance creative control in photographic applications, allowing users to correct focus errors or achieve artistic effects that were previously only possible during capture.

Key insights

AI refocusing allows dynamic post-capture adjustment of photographic focus and depth of field from a single image.

Principles

• Defocus blur is determined by the circle of confusion.
• Depth maps enable precise defocus map computation.
• Pre-trained models can be fine-tuned for realistic blur.

Method

The method involves estimating depth and defocus maps from an input image, then training two models: one to deblur to an all-in-focus image, and another to apply desired defocus blur using a pre-trained text-to-image model.

In practice

• Adjust aperture size after photo capture.
• Shift focus plane to different subjects.
• Sharpen all subjects in group photos.

Topics

• AI Image Refocusing
• Computational Photography
• Defocus Blur
• Depth Estimation
• Generative AI Models

Best for: AI Engineer, Machine Learning Engineer, AI Student

Related on AIssential

• It's all about the angle: Your photos, re-composed — A new Google Photos feature uses ML and generative AI to re-compose 2D images as 3D scenes, enabling perspective changes.
• 🎯Generative Refocusing is out🎯 👉Generative Refocusing is a two-step process that uses... — Generative Refocusing uses a two-step neural network process for deblurring and controllable bokeh generation.
• Evo-Depth: A Lightweight Depth-Enhanced Vision-Language-Action Model — Evo-Depth enhances VLA models with implicit depth encoding from RGB images for improved spatial understanding in robotics.
• Direct 3D-Aware Object Insertion via Decomposed Visual Proxies — DIRECT enables 3D-aware object insertion by decomposing guidance into appearance, geometry, and context, injected separately for precise control.
• VDPP: Video Depth Post-Processing for Speed and Scalability — VDPP improves video depth estimation post-processing through geometric refinement, offering speed, accuracy, and scalability.
• Training the AIs' Eyes: How Roboflow is Making the Real World Programmable, with CEO Joseph Nelson — Computer vision, though trailing language models, is approaching its "ChatGPT moment" through specialized, efficient, and edge-deployable models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Jia-Bin Huang.