Images used in video and motion graphics projects rarely arrive in exactly the right format. A strong photograph might be cropped too tightly for a 16:9 frame. A client's only available image of a subject might be missing the visual space needed to work in an animation context. An image sourced for reference might be the wrong aspect ratio, the wrong orientation, or simply cut off at an inconvenient place. These are common, recurring problems in production, and they historically required either finding a better image or spending significant time in Photoshop with cloning and retouching tools.
- Generative expand lets you extend an image beyond its original borders using AI to fill in the new areas. This is especially useful for converting portrait or square images into the 16:9 landscape format standard in video work.
- Generative fill allows you to make a selection within an image and have AI generate new content within that selection, either from a text prompt or simply by analyzing the surrounding image and filling contextually.
- Neither feature produces perfect results on the first attempt in every case. The practical approach is iterative: generate, evaluate, try different selections or models if needed, and combine AI-generated content with traditional Photoshop tools to finalize the result.
This lesson is a preview from our Generative AI Certificate Online. Enroll in a course for detailed lessons, live instructor support, and project-based training.
This is a lesson preview only. For the full lesson, purchase the course here.
These features are available in Photoshop's Creative Cloud releases, including both the standard Photoshop application and the public beta version, which occasionally includes capabilities before they arrive in the main release. If you have a Creative Cloud subscription, you can access the beta through the Creative Cloud desktop app. Standard and premium subscriptions differ in whether you have access to partner AI models beyond Adobe's own Firefly system.
Setting Up an Image for Video Format
One of the most common situations where generative expand is useful is converting an image into a 16:9 aspect ratio. Video content runs at 16:9, which is a landscape format. Many available images, particularly photographs taken for web or print use, are in different proportions. A cropped editorial photograph might cut off important content at the edges. A vertically oriented image will not work at all without significant changes. The crop tool in Photoshop can extend an image's canvas rather than simply cutting it down, and this is the starting point for using generative expand effectively.
Open the image and convert the background layer into a standard editable layer if it is locked. Then select the crop tool and choose a 16:9 preset from the aspect ratio options. Instead of pulling the crop boundary inward to cut away parts of the image, extend it outward to create space around the existing content. Holding the Option key on a Mac, or the Alt key on Windows, while dragging a crop handle extends the canvas symmetrically on both sides. Apply the crop, and you will have your image surrounded by empty canvas space in the areas where new content needs to be generated.
You can make the canvas larger than you strictly need at this stage. It is easier to trim back a little extra space later than to realize you did not leave enough room for the AI to generate convincingly. Having a bit of overlap between the existing image content and the selection you use for generation also helps the tool understand what it is supposed to be extending, which generally produces better results.
Running Generative Fill Without a Prompt
Once you have the empty canvas areas set up, use the rectangular marquee tool to make a selection over one of the blank regions you want to fill. The contextual task bar that appears near your selection will include a generate button. In current versions of Photoshop, this opens the generative fill options where you can choose your AI model and, optionally, write a text prompt describing what you want the fill to produce.
For many image extension tasks, leaving the prompt field blank and letting the tool analyze the surrounding content produces the most coherent results. Photoshop's Firefly fill and expand model is specifically designed for this kind of contextual image extension, and it often does better without a text prompt than with one. Without a prompt, it reads the surrounding pixels and tries to generate something that matches the existing content as naturally as possible. With a prompt, it tries to balance the surrounding context with your description, which can sometimes lead to results that feel disconnected from the rest of the image.
Each generation gives you three options to choose from, visible in the properties panel. Review all three before selecting one. The differences between them can be subtle or significant, depending on the content, and what looks best is not always obvious from a thumbnail. Zoom in on areas where the generated content meets the original image to evaluate how the seams look, since this is where problems most commonly appear.
Working in Sections Rather Than Making One Large Selection
For images where multiple areas need to be extended or filled, working in sections generally produces better results than trying to fill everything with a single large selection. A large empty area gives the AI a harder problem: it has to generate more content with less surrounding context to analyze. Filling areas in smaller pieces, each with more overlap with the existing image, tends to give the tool better information to work from and tends to produce results that integrate more smoothly.
For a typical image-to-16:9 conversion where content needs to be added on both sides and possibly above or below, a reasonable approach is to fill the top or bottom area first, then fill each side separately, then address any overlap areas or seams that remain. Each generation is an independent operation that adds a new layer to the Photoshop document, so you retain the ability to go back and choose a different variation or regenerate a section if you are not satisfied with the result.
The key is to look at each generated section critically before moving on. Sometimes the first generation is excellent. Sometimes it is close but not right. Sometimes it produces something completely unusable. In all three cases, the response is the same: evaluate what you got, decide whether to accept it, generate more variations if you want to, or try a different approach if this one is not working.
Using Different Models and Prompts to Get Better Results
Photoshop offers access to multiple AI models for generative fill. The default Firefly fill and expand model is the most reliably useful for image extension tasks. Beyond that, Adobe offers partner models from third-party AI companies, accessible to users with premium subscriptions. These partner models consume premium credits rather than standard generation credits, so they are a more limited resource, but they can produce significantly different results that are sometimes more useful for specific types of content.
When the default model is not producing what you need, it is worth trying a different one. The range of available partner models changes as Adobe develops these integrations, but in practice, the differences between them can be quite significant. One model might handle the texture and color of a natural environment better than another. A different model might do a more convincing job with specific types of subjects. There is no reliable formula for which model works best in a given situation; it requires experimentation.
Adding a text prompt can also change the results. If the contextual analysis alone is not producing what you need, describing the content you want, whether that is the background environment, the specific element you are trying to extend, or the general visual quality of the fill, can push the generation in a more useful direction. The prompt does not guarantee that result, but it gives the model more information to work with.
- Try the default fill and expand option first before moving to partner models, since it is designed for extension tasks and tends to integrate most naturally with surrounding content.
- When adding a text prompt, describe what the fill should contain rather than what style it should have. Spatial and content descriptions work better than aesthetic descriptions for this type of task.
- If a partner model produces something much more or less detailed than the surrounding image, the seam between generated and original content will be more visible and harder to clean up in post.
What to Do When the AI Result Is Close but Not Quite Right
A common outcome with generative fill and expand is output that is mostly what you need but has a specific problem: an edge that does not quite blend, a generated element that is slightly the wrong size or shape, a texture that is close to the surrounding image but not quite consistent. These are situations where the right approach is to accept the AI-generated content as the working material and use traditional Photoshop tools to fix the specific problem rather than continuing to generate new versions.
The spot healing brush is particularly useful for blending seams and fixing areas where generated content does not quite connect with the existing image. The clone stamp tool lets you replicate specific areas of texture or content from one part of the image to another, which is useful when generated content is in the right area but does not match the visual pattern of what surrounds it. Standard retouching techniques that work for any compositing task will work here, too, because the generated content, once you accept a variation, is just pixels like any other.
Once you have worked with both AI-generated layers and traditional tools to get the image to a good state, merging layers is a reasonable step before doing final cleanup, since it gives you a unified surface to work on. Note that merging breaks the connection to the generative layers, so you lose the ability to switch between variations after that point. Make sure you are satisfied with the overall direction before merging.
Saving and Using the Extended Image in Your Project
When you are satisfied with the result, save the file in whatever format your production workflow requires. If you need to preserve the generative layers for further adjustment, save as a Photoshop document. If the image is going directly into a video or motion graphics composition, exporting as a JPEG or PNG is the most common approach, with the choice between them depending on whether you need transparency.
The extended image can then be brought into your motion graphics or video project as you would any still image. In many production contexts, a still image extended to a 16:9 format is used as background footage, as a subject for a Ken Burns-style pan or zoom animation, or as a compositional element layered with other materials. The generative fill and expand process gives you the spatial room to work with the image in these ways, even when the original was not sized or framed for video use.
Keep in mind that the standard for what is acceptable depends on the production context and the audience. For a general commercial project where the image is one element among many and the viewer is not expected to examine it critically, AI-extended content that reads naturally at normal viewing distance is entirely usable. For a production where a subject matter expert or a highly attentive audience will be looking closely at specific details, the bar is higher, and the AI-generated content will need to be scrutinized more carefully before it is acceptable.
How to Develop a Reliable Iterative Process
The most important thing to understand about using generative fill and expand effectively is that it is an iterative process, not a one-step solution. On any given image, the first generation might be excellent, or it might require several rounds of attempting different selections, different models, or different prompt approaches before you get something you can work with. Building your expectation around iteration rather than immediate success makes the process significantly less frustrating and more productive.
The pattern that works reliably is: generate and evaluate, adjust your approach based on what you see, generate again, and repeat until you have something close enough to finish with traditional tools. What changes between iterations might be the size or position of your selection, the model you are using, whether or not you include a text prompt, or the number of times you generate similar variations before accepting the best result from a given set.
Over time, you develop intuitions about which types of images and which types of extension tasks respond well to which approaches. Natural environments and photographic backgrounds tend to extend reasonably well. Subjects with specific anatomical or structural logic, like animals or people, often require more work to get right. Images with complex patterns or textures at the edges of the extension area will be harder to match convincingly than those with simpler, more uniform surroundings. These intuitions take time to develop, but make the process considerably more efficient once they are in place.
