sharp gull
Sometimes you may need to create something more complex than a single 1024x1024 pixel image, which is the standard output from DALL-E 2. You may also have certain requirements for the task that limit what you can create. In these cases, you may need to follow a process that includes image generation as part of an editing workflow.
Full tutorials on how to use the Dall-E 2 editing tool are available in #📚┃resources. We will show a simple example here to help you get started. This workflow can be applied to many similar tasks.
For example, let’s say you need to make a banner for a Discord channel using only the Dall-e editor and a cropping tool, which can be found in most image editing software. For banners, Discord will let you upload an image that is 1920x1080 landscape format. You can use Dall-E 2 to create a base image in 1024x1024 square format, then use the Dall-E 2 editor to increase the size of your image by adding frames.
In this example, we will use an original image and additional frames to expand the canvas both horizontally and vertically. Mathematically, four frames are the minimum number needed to reach a width of at least 1920 pixels, as we will need to add at least 896 pixels horizontally. This is about 87.5% of a 1024x1024 image, which would give about 10% overlap between the new frame and the base image to maintain style consistency. We also need to increase the vertical height from 1024 to 1080.
The larger the canvas you create, the more options you will have for cropping and adjusting the final image. Keep in mind that the size of the canvas will depend on your specific needs and requirements.
To create a consistent art style when adding frames, you can use a consistent prompt format such as "[subject][style]" for all of the frames you generate. This will allow you to keep the same art style cues, but vary the subject being rendered. Maintaining a consistent style across different subjects can be challenging, as some styles may work better with certain subjects. To improve the chances of consistency, you can choose a style that produces consistent results across a full set of images.
Here is “a winter landscape of ice skating on a frozen pond. view from a distance. by norman rockwell and matisse.”
These four images have a similar style, which suggests that our art style cues are producing consistent results. We can choose one of these images and use it as a base to increase the size of the canvas through outpainting. Each of these images is 1024x1024 pixels.
To start editing, select the image you want to use as a base, then click the "edit" button.
Let’s add generation frames on the right and left sides, with enough overlap so the art style can be matched well. Let’s keep the prompt unchanged. You should also make sure to include a good amount of overlap between each added frame to help the AI maintain style continuity. The more overlap we have, the better the chances of maintaining style consistency. For this example we will add two frames horizontally, allowing us to keep good overlap between them.
To increase the vertical dimensions of the image, we can add frames above the base image, making sure to leave enough overlap to maintain style continuity. When adding frames, it may not be necessary to mention the skaters in the prompt because the part of the image we are extending is primarily focused on the background landscape. We can simply use a prompt that describes the style and composition we want to achieve in the added upper frames.
For the next frame we will drop the mention of the ice skating because we are focused on the upper landscape. Every word in a prompt affects the results, and omitting unnecessary goals usually helps coherence.
As you add frames to the image, you may want to make changes to certain parts of the previous frame. To do this, you can erase the parts you want to change, and the AI will regenerate those parts of the frame using the current prompt. For example, here we will erase the tall trees. The AI will generate new content based on the updated prompt. This allows you to make changes and adjustments to the image as you work to create the final banner.
Now we can download the edited image by using the download arrow.
The download icon is in the upper right in this example.
This image is now 2496x1536, larger than the 1920x1080 size we need for a Discord banner.
Many cropping tools allow you to specify a specific crop size, such as the 1920x1080 pixels needed for a Discord banner. You can preview crops of different parts of the canvas to see how they would look as the final banner. This can be helpful for visualizing different crop options and deciding on the best composition for your banner. By using a cropping tool, you can fine-tune the final composition to ensure that it meets your specific needs and requirements. Here are visualizations of how three different crops could be made from our large canvas
Here is an example of a crop from our main canvas that avoids placing a large key element in the center. The foreground skaters would probably catch the eye in the final version.
When creating a banner for Discord, keep in mind that the server name may appear as text on top of the banner, so you should leave the top 48 pixels free of any busy design. Additionally, you may need to consider other design constraints, such as the need to add text in the center of the banner, which will cover any central image. Here are two examples of how composition would be affected by placement of a text overlay in the center of an image. Dall-E 2 is not designed to produce text, so you will need to use a separate image editing tool if you want to include text in your banner.
