DALL-E 3 Now Has Integrated Editing Tools, But They Need Work

Ever wanted to edit your DALL-E 3 creation immediately but did not know how to make the changes?

OpenAI hears you and has added inline editing tools to DALL-E 3 in ChatGPT, allowing you to make tweaks to your creations without leaving the app.

The only issue is that they don’t do what OpenAI advertises, and it’s a mixed bag as to whether you’ll get the edits you want. But still, it’s a start of what’s to come for DALL-E 3 editing, and this is how they work.

They’ve been a while coming, but OpenAI has finally added editing functionality to DALL-E 3 in ChatGPT Plus. Prior to this, you could attempt to make edits to your DALL-E 3 creations, but it would typically always result in a completely fresh image being generated. Even if the prompts used to create the edited image were extremely similar to the first prompt, due to the very nature of generative AI, the output could be significantly different.

DALL-E 3’s integrated editing tools now aim to fix issues within the same image without creating a new prompt. You highlight an area in the image you want to adjust, prompt DALL-E to make a specific change (such as “remove this” or “add this feature”), and it goes to work.

However, it’s a mixed bag.

OpenAI’s official blog example shows DALL-E adding cherry blossoms to a Ukiyo-e-style image of cat, and it working perfectly. In my experience, DALL-E’s editing tools struggle to add such detailed imagery to an existing image and, for the most part, can’t process requests of this nature.

That’s not to say the editing tools don’t work. They can absolutely remove features and make small changes, as detailed below, but it’s all on a small scale.

The best way to show you what DALL-E 3’s editing tool can do is with some proper examples.

Removing and Erasing

DALL-E 3’s editing tools can erase some small features, but it definitely struggles. In the following image, I’d asked DALL-E to create “A robotic hand holding a glowing, transparent square chip with ‘AI Accelerated’ written on it. The background is dark blue, creating a futuristic and high-tech atmosphere.”

Now, given writing text is a specific issue for generative AI, I’m not surprised the text was wrong—so I asked DALL-E to remove the spelling mistake. It failed on the first attempt with the prompt “remove this please” but succeeded on the second try with the prompt “please remove this word from the image.” So, success, yes, but you may have to try different prompts to get a result. I then asked it to move the AI letters to the center of the square, which it did well despite crashing my PC for some reason.

Next up, I asked DALL-E to help me edit an image of a television showing static. DALL-E is great for generating unique images, but it gets some details wrong or delivers warped text. In this case, the information bar on the TV screen displays gibberish AI text, so I wanted to remove it. DALL-E’s first attempt left a shadow where the information bar was, while the second and third tries were much cleaner. Not perfect, but decent enough for what I was using the image for.

Adding and Changing

DALL-E 3 can make some changes to images, but again, it depends on the scale of what you’re asking. Small additions and changes are possible, but most of the time, DALL-E’s editing tool can’t process what is asked.

I asked DALL-E to “create a 2:1 4k image of a human eye close up, be detailed,” resulting in the following image:

Pretty decent, if a little odd (eyelash reflections?!), but I wanted to switch the iris color to blue-green instead, resulting in this image:

The color is slightly dull but unchanged, and some features are warped slightly.

Next up, I prompted DALL-E to “create a 4k 2:1 claymation style image of a person wearing a tiger onesie skiing down a slope with a yeti background.”

Again, the original image is good, has sharp, vibrant colors, and matches the prompt. However, I wanted to switch the yeti to a different type of scary animal.

Instead of swapping the yeti for a different animal, it’s just completely gone. Instead, I returned to the original image and asked it to add another yeti, which it failed to do. It similarly failed when asked to add more trees to the image.

Adding Text

There isn’t much to cover here. As said, generative AI generally struggles with text anyway, and adding text to an image is even more difficult. I prompted DALL-E to create a postcard from a British seaside, which it did, capturing the scene nicely. I then wanted to add the words “Sunny Beach” to the image, which it just couldn’t handle. That is, until after around six or seven back-and-forth prompts reminding DALL-E what it needed to do, it created a whole new image with the required text. And accurate text, too; I’m not sure about the baby face in the sunshine, mind.

Next up, I went for a birthday card. This time around, DALL-E added “Happy Birthday” without being prompted, which is fine, given the theme. So, I asked it to add the date of the birthday on the card. Again, it failed to add the date to the card, but after a few rounds of prompting, it created a new image featuring the requested date. In this case, adding the date myself would have been easier, especially if I wanted to keep the original image.

I’ve found that DALL-E’s editing tools are useful for certain tasks, but it takes some time to get the editing right. At other times, it just won’t work, and you’re probably better off taking your original prompt and tweaking that to create a new image.

DALL-E’s editing tools will likely improve over time, but until then, you might be better off editing your image with an image editing tool rather than going through multiple time-consuming rounds of conversation with DALL-E. When it works, it’s good; when it doesn’t understand or doesn’t make the edit at all, it’s frustrating.