What's for breakfast? Generating unappetizing food images

The idea

One morning I had a bland, burnt egg and half a croissant that didn’t survive the toaster oven. The day before, I’d been working on image generation for volunteering — prompting a burger and a glass of pop overlooking a scenic landscape. I was struggling with camera angles and perspectives.

Then I thought, maybe the real challenge isn’t making food look good—it’s capturing something that looks as unappetizing as my breakfast.

What's for breakfast?

Try it out

Heads up!

There is some unpleasant food imagery, generated with AI.

Prompting for consistency

ChatGPT gave me a list of 20 common and sad breakfast ideas, from burnt eggs, to pop tart in water and melted ice cream. I knew I wanted to generate over 10 images and try it all. I also knew I wanted them to look consistent: light, angle, relative size, position.

This prompt helped me with achieving consistent results and save time on editing:

💡

Create image. Studio light, white background, colour graded. Fit the object into the view.

Gemini's love language is ✨ perfection ✨

Unless I specified the unappetizing details, the model kept generating perfect-looking food. Words like "second day," "dry," or "sad" didn’t really register with Gemini. So I started building a vocabulary of what actually makes food look unappealing. I had just canceled my ChatGPT subscription, so I was sticking it out with Gemini for now.

A toast with cucumbers, a roll of a cheese slice, rice with Doritos and a drizzle of hot sauce.

Prompting for imperfection

Who hasn’t had eggs for breakfast? I’ve had more than my share of overcooked ones—despite liking them sunny side up and runny. Funny how I assumed my version of “unappetizing” was universal. (Hello, bias!) Turns out, I had to be pretty specific about how I wanted my egg to look burnt.

💡

Create image. Studio light, white background, fit the object into the view. Over easy pan fried egg, overcooked, edges are overly burnt. The yoke is broken. Add a few hairs, the plate is dirty

Taking prompts too literally

Describing things with context can be a fun and unexpected way to generate images, and it often leads to some very interesting and surprising results.

💡

... Create an image of dry ramen noodles crushed over a bowl of hot steamy water. The are flavour packets on the side

Prompt below put the chocolate bar literally under the bed.

💡

... Create an image of a protein bar someone dropped under the bed last week, it's half eaten and is rolled in dust, the packaging is ripped and scrunched.

It’s hard to be mad when the model gives you something totally unexpected—especially when it somehow does match your prompt and makes you laugh in the process.

Unappetizing fast food is a challenge

Creating an image of unappetizing fast food is surprisingly difficult. For one, it’s designed to look appealing—bright colors, perfect proportions, glossy textures. And beyond that, our brains are wired to associate burgers and fries with indulgence, comfort, and satisfaction. It takes some serious prompting to break that illusion.

It wasn’t until a few generations in that I realized the prompt was actually encouraging proper lighting and composition—things that naturally enhance a food’s texture and color. To get the result I wanted, I had to tweak the prompt to intentionally make the food look dull or unappealing—without making it look like a bad image overall.

What about switching models?

Turns out, ChatGPT with DALL·E is surprisingly good at personifying feelings in objects. I only had to prompt once, and this is by far the saddest image of greens and fast food I’ve managed to get. Just look at that salad face—pure veggie despair.

Composition and fitting objects into the view

After countless tries, I still couldn’t figure out the right prompt to get all the objects to fit into view. Most of the time, Gemini insisted on placing forks and knives awkwardly outside the frame. Even adding “fit all objects into view” didn’t seem to help. And since Gemini doesn’t support incremental changes, I had to keep resending the same prompt (sometimes out of pure stubbornness) only to get a completely new image each time.

Issue persisted with a spoon of peanut butter

Composition turned into a real challenge when I tried to create an image of a spoon with peanut butter on it. After some trial and error—mostly tweaking the word order—I finally got lucky and managed to get the spoon fully in view:

💡

.. The spoon is laying on the angle with some crunchy peanut butter. The spoon is in the view, I should be able to see the handle.

Controlling liquid level

It's ether full, or empty, I could not get Gemini to do 1/3 of a container of any liquid kind.

Pizza crust

After over 12 tries, Gemini still couldn’t manage to generate a pizza crust! I even provided references, but no luck. I didn’t realize I was capable of describing leftover pizza crust in so many details.

💡

Create image. Studio light, white background fit the object into the view. Eaten pizza crust with a bit sauce on it, there are bites and no toppings left, sitting on a broken greasy pizza box.

Giving image reference to Gemini to create a pizza crust with bite marks, and it completely ignores the request.

Discovering appropriate prompting and slowly getting there with ChatGPT DALL-E!

ChatGPT DALL-E knew that crust was, it was failing to recreate bite marks, so I asked it for exactly that, then sprinkled some toppings on it.

Studio lightning will make things look better

I was leading my prompts with “studio light, white background” to keep things consistent. Even though the food still looked dull, the images remained relatively positive. The first two images were generated by Gemini, then fed into ChatGPT. While the composition and groceries stayed the same, it was the lighting that really shifted the mood of the image.

Editing images

Consistency was the goal since the beginning of the project, so I picked images that looked well together. However, many of them required background removal and some colour adjustment.

Photoshopping with florafauna.ai

I was curious if AI would also be able to photoshop my images. Flora is a web tool for combining imagery using the same AI models. Unfortunately, the background removal was too rigid. It would remove objects entirely, or overexpose the image. Maybe I wasn’t prompting it right, as it didn’t quite work as I’d hoped.

At this point I wanted wrap this project up, a quick background removal in a familiar tool and a mask helped me with achieving consistency across images.

Conclusion: Gemini VS ChatGpt DALL-E

💡

Prompt: White background, studio light, colour corrected, very sad second day crushed croissant from a grocery store.

I believe these two images perfectly summarize the differences in generating unappetizing food. Gemini Flash can create stunning, photorealistic images in high quality within seconds. On the other hand, ChatGPT DALL-E takes about a minute to produce a lower-resolution, high-noise image, yet a very creative oddly crushed croissant.

Gemini

✅ Free!

✅ 2-5 second image generation when available

✅ Photorealistic

❌ Inconsistent image re-generation when correcting small details

❌ 1:1 aspect ratio

ChatGPT DALL-E

✅ Consistent image re-generation when correcting small details

✅ Customizable aspect ratio

✅ More creative with generating imperfections

❌ Much lower image quality, high noise in poor illuminated images

❌ Limited free image generation

If you need imperfection in your imagery I would highly recommend going along with DALL-E if you're willing to pay. Else buckle up to discover millions of way to describe the imperfection you're imagining with Gemini.

A few images that did not make it

Noodles on a bill, leaky banana standing up right (why?), and dried cucumbers on a moldy toast.