JD.com Launches JoyAI-Image-Edit, Entering the AIGC Image Race with “Spatial Intelligence”
Recently, JD.com has made another major move in the AI field by launching and open-sourcing JoyAI-Image-Edit, an integrated image model. The model’s core innovation lies in deeply integrating “Spatial Intelligence” into the image understanding and editing process. The goal is for AI to move beyond 2D pixel manipulation to truly understand and operate on spatial relationships in the 3D physical world.
This move is seen as a concrete implementation of JD.com’s AI strategy: leveraging its core supply chain advantages to extend technology into the physical world and real-world industries, particularly in key areas like e-commerce content creation and embodied AI training.
Spatial Intelligence: A Technological Leap from “Photoshopping” to “Manipulating Space”
Traditional image editing models often fall short when handling spatial relationships. For instance, when moving or replacing objects, issues like perspective errors, inconsistent scaling, and mismatched lighting and shadows are common. The root cause is the models’ lack of a deep understanding of 3D geometric space.
JoyAI-Image-Edit addresses this pain point with a systematic design. Its main features include:
- Core Capabilities: On top of supporting over 15 general editing tasks, the model features “spatial editing” as a core capability. It supports spatial operations on objects within an image, such as moving, rotating, and changing viewpoints. It can also understand commands with precise geometric parameters like “move left by 0.3 meters” or “rotate 45 degrees,” achieving a high degree of controllability in the editing process.
- Technical Architecture: It uses a unified architecture combining a Multimodal Large Language Model (MLLM), a Variational Autoencoder (VAE), and a Diffusion Model (MMDiT). The MLLM is responsible for high-level semantic and spatial relationship understanding, while the diffusion model executes precise generation and editing tasks based on this understanding.
- Data-Driven: To train the model’s spatial awareness, JD.com’s research team built a new data system, including the OpenSpatial-3M dataset with 3 million images. This dataset, created using multi-view synthetic data and spatial editing data with precise pose parameters, guides the model to learn real-world geometric principles during training.
Performance Benchmarks: Reaching the International Cutting Edge on Multiple Metrics
According to public benchmark data, JoyAI-Image-Edit has demonstrated outstanding performance in various standard tests, placing it in the top tier internationally.
- Spatial Understanding: In 13 benchmarks covering 2D semantic perception, 3D spatial understanding, and 4D spatio-temporal reasoning, the model showed significant improvements in 9 spatial understanding-related benchmarks, achieving an average score of 64.4, on par with the closed-source model Gemini 2.5 Pro.
- Spatial Editing Precision: On the SpatialEdit-Bench, designed specifically for spatial editing capabilities, JoyAI-Image-Edit scored an Object Overall Score of 0.649 and a Camera Overall Score of 0.571. This not only surpasses existing image editing models by a large margin but also exceeds the spatial editing precision of some video world models like Veo3.1, ViduQ2-Turbo, and Kling.
- General Editing SOTA: On GEdit (focusing on Chinese instructions and real user needs) and ImgEdit (focusing on comprehensive capabilities and reasoning), two authoritative industry leaderboards, JoyAI-Image-Edit achieved scores of 8.27 and 4.57 respectively, setting new SOTA (State-of-the-Art) records for open-source image editing models.
Core Application Scenarios: Empowering E-commerce and Embodied AI
The powerful spatial capabilities of JoyAI-Image-Edit make it highly valuable in two areas closely related to the physical world.
- E-commerce Content Creation: In e-commerce, high-quality, multi-angle product images are key to attracting customers. With JoyAI-Image-Edit, merchants can avoid expensive and time-consuming photoshoots. For example, based on a single product image, they can use commands like “rotate the shoe to a front-facing view” or “adjust the sofa’s position in the living room” to quickly generate a series of multi-angle, multi-scene assets that adhere to spatial logic. This maintains natural consistency in lighting and proportions, significantly reducing content creation costs.
- Embodied AI Training: The development of embodied AI robots heavily relies on vast amounts of high-quality real-world data, but data collection is costly. JoyAI-Image-Edit can generate high-quality synthetic data with spatial consistency to augment training datasets. This approach can effectively assist robots in learning spatial reasoning, for instance, by generating novel view images (Thinking with Novel Views) to help them better “see” and understand their physical environment, thus accelerating their training and iteration process.
JD.com’s AI Strategy: Open-Sourcing and Industrial Application in Tandem
The open-sourcing of JoyAI-Image-Edit is another significant step in JD.com’s AI strategy. Combined with its recently open-sourced JoyAI-LLM Flash model, the ongoing construction of the world’s largest embodied AI data collection center, and the JoyInside platform for embedding AI capabilities into hardware terminals, it’s clear that JD.com is building a complete closed loop from models and data to terminal applications.
Its strategic path is clear: on one hand, lower the technical barrier through open-sourcing to activate the developer ecosystem; on the other hand, deeply integrate AI capabilities into its foundational business scenarios like supply chain, logistics, and retail. This lets the technology create value and iterate by solving real business problems. This pragmatic approach aims to transform AI into a new engine driving its business growth.
