.svg)
.png)
Realistic and high-impact brand images are crucial to connecting with customer audiences. When visuals look authentic, a customer can truly envision themselves engaging with the product or experience.
Unfortunately, images like this typically come with high costs and time-intensive photoshoots, making it challenging for brands to adapt quickly and change out imagery alongside new branding, campaigns, regions, or seasons.
As a Research Scientist on the Jasper Research team, I'm excited to share that we’ve introduced a new single-step model that predicts and controls shadows for objects in images, as well as a test set and a demo. We’ve spent the last couple of months working on this cutting-edge tool, which delivers remarkable speed, while offering precise control over the direction, softness, and intensity of the generated shadows, enabling unparalleled customization.
For marketing specifically, our shadow generation model allows marketers to create lifelike and fully-customizable shadows for their images in seconds—no expensive setups, no drawn-out editing processes. By bridging the gap between realism and efficiency, Jasper is giving brands the power to craft visuals with greater agility and within their budgets. Want to explore Jasper's image capabilities? Learn more about our AI Image Suite here.
What Is Shadow Generation and Why Does It Matter to Marketers?
Shadows are more than a byproduct of lighting—they’re a vital element in creating realistic and visually engaging imagery. A well-placed shadow adds depth, texture, and dimension, transforming flat visuals into lifelike images. This is especially critical in fields like e-commerce and product photography, where customers rely on images to gauge product quality and appeal.
Examples below:
Though it may seem like a small detail, it has a significant impact on how consumers perceive your products. Consider your experiences with online shopping. Products that appear in realistic settings—like the ones where you’ll actually use them—likely grab your attention more quickly and give you more confidence in your purchase.
Customers may take these images for granted, but creating them is often expensive and time-consuming. It involves photographers, set designers, equipment, and lots of time taking and editing photos.
For brands with tight budgets or high content demands, these barriers can result in compromises— either settling for lower-quality visuals or overspending on resources. Jasper's shadow generation model changes the game entirely.
By combining cutting-edge diffusion models with intuitive controls, it lets marketers create realistic, customizable shadows in seconds.
Users can tweak shadow direction, softness, and intensity as easily as adjusting a filter. Want a softer shadow that blends naturally into the background? Done. Need sharper edges to emphasize bold product lines? No problem. The tool adapts to your creative vision with a few simple adjustments, putting full creative control back in your hands.
The video below shows you how. In the top left corner of each video, we illustrate the light source using a circle within a spherical coordinate system. This display includes the polar angle, azimuthal angle, light size parameter (which controls shadow softness), and the shadow intensity parameter, offering a clear visualization of how these factors influence the generated shadows.
This level of precision and efficiency means brands no longer need to sacrifice quality for speed—or break the bank to achieve professional results. Instead, they can produce polished, on-brand visuals at scale, ensuring every image resonates with its intended audience.
How We Train the Model
Shadows aren’t simple shapes—they depend on light sources, object geometry, and context, which makes creating a training dataset particularly complex. Manually generating shadows for training is not only time-consuming but also costly, requiring professional annotators to craft shadows that are both geometrically precise and visually accurate.
To overcome these challenges, I took an innovative approach: using a rendering engine to generate a fully synthetic dataset.
Building a Synthetic Dataset
The first step was to curate a diverse collection of 3D meshes—digital representations of real-world objects. These meshes, created by professional artists under free-use licenses, represented a broad range of shapes, sizes, and materials, ensuring the dataset could handle a variety of use cases.
With the 3D objects ready, the rendering engine simulated lighting conditions by randomizing light source parameters. These included:
- Polar and azimuthal angles: Controlling the direction of the light source
- Light size: Adjusting shadow softness
- Shadow intensity: Modifying how dark or light the shadows appear
By iterating through these parameters, we rendered thousands of images with different lighting setups, capturing variations in shadow direction, sharpness, and depth. Each image included the object, its mask, and a shadow map, creating a dataset rich in diversity and precision.
Training with Light Parameters
To give the AI precise control over shadow generation, the model was trained to understand light parameters directly. By incorporating data like polar angles and light intensity into the training process, the AI learned to map these inputs to realistic shadow outputs.
We also utilized rectified flow, a technique that optimizes the model’s ability to predict shadows in a single step. Unlike traditional methods that rely on multiple stages, this streamlined approach enables faster inference without sacrificing accuracy.
Visualizing the Shadow Prediction Process
We use 3D meshes, a camera, and a light source to build a dataset by rendering images using a rendering engine. This setup allows for dynamic lighting variations and ensures shadows behave naturally across a range of conditions. Then we use this dataset to train our shadow prediction model, which takes an object image and the light parameters as inputs. The result? A model that can generate realistic shadows for any object in seconds, while giving users full control over their customization.
The figures below show the spherical coordinate system and complete shadow generation pipeline (Figure 3):
Open Source Contributions
When developing Jasper’s shadow generation model, we encountered a challenge: There wasn’t an existing dataset to evaluate the tool’s performance. Rather than letting that limitation slow progress, we created our own benchmark dataset tailored to this unique task—and we’re sharing it with the research community.
This benchmark dataset sets a new standard for AI research in shadow generation. It includes three distinct tracks designed to test a model’s ability to control shadow characteristics like softness, horizontal direction, and vertical direction.
Each track evaluates how well a model can customize shadows based on these parameters, providing a clear framework for comparing and improving AI tools. You can access it here.
By releasing this dataset as open source, Jasper aims to foster collaboration and innovation across the AI community and support the development of even more advanced tools.
A New Era for Visual Content
Jasper’s shadow generation model is a glimpse into the future of visual content creation. By combining realism, speed, and customization, it’s reshaping how businesses handle product photography and marketing.
Potential applications are diverse and exciting. In e-commerce, realistic shadows can make product visuals feel more grounded and appealing, helping customers trust what they see. Imagine augmented reality applications where customizable shadows enhance how virtual objects blend into real-world settings. Beyond marketing, industries like film, gaming, and virtual reality could use this technology to create immersive environments with dynamic lighting and realistic shadow interactions.
As technology continues to advance, we are excited to further redefine the creative process, making it easier to produce visuals that are both professional-grade and versatile enough to keep pace with shifting trends and demands.
Want to play around with the model? Take a spin with this shadow generation demo.
Plus, explore our research on Flash Diffusion—a new method designed to improve the efficiency and speed of diffusion models for AI-driven image generation. We’re excited to share that this paper has been accepted by the Association for the Advancement of Artificial Intelligence (AAAI)!
.png)
