Generative AI tools: Is Freepik new Reimagine any good?

Last Modified: May 6th, 2024 - Category: Artificial Intelligence, News, Web Design
Reimagine generative AI tool cover image

The Spanish company Freepik has launched its new generative AI tool called Reimagine, and we tested it to see how effective it is.

If you have read our previous four-part case study comparing various generative AIs for images and videos, you might feel that new tools are emerging almost every day. Indeed, it’s challenging to keep up with the burgeoning number of new generative tools based on artificial intelligence, particularly for multimedia—images, videos, and audio.

However, in the case of Freepik, I believe it deserves to be added to the list of tools mentioned in the previously referenced case study.

The reason for this inclusion is Freepik’s status as the “de facto” leader in graphic resources distribution for web designers, webmasters, and UI design. Therefore, the announcement of its own AI-based image generation tool is significant and warrants consideration.

Freepik Reimagine AI Tool Analysis

If you have read the aforementioned case study on generative AI tools, you know we conducted an extensive study using UX research techniques. The study was long and complex, and the results were quite precise.

However, for this test, I do not have the luxury to undertake that entire process, so I will perform an analysis based on a few original images and how Freepik’s Reimagine works with them.

List of AI models and types

The list of preset models is really cool, and there are many well-trained AI models to choose from. Each model has three options: Subtle, Vivid and Wild.

Here is the full list of generative AI models:

  • No style
  • All styles
  • photo
  • cartoon
  • 3d
  • vector
  • vintage
  • digital art
  • comic
  • painting
  • dark
  • sketch
  • mockup
  • 2000s phone
  • cyberpunk
  • 70s vibe
  • anime
  • watercolor
  • studio shot
  • art nouveau
  • pixel art
  • origami
  • surreal
  • fantasy
  • traditional japan

First AI Generation Test: Rimbaud

For our first test, I’ll use an image of the French poet Arthur Rimbaud. The image is an oil on canvas painting of Rimbaud based on an old black-and-white photograph. I am not sure who did this recreation (and it might be an AI generation as well), but I think it has the appropriate level of detail for an initial test without being overly complicated.

One thing to mention: there are many preset models, so we’ll test just a few to demonstrate how it works. Testing all models would be burdensome, but you can test them yourself if interested.

Original image:

Image of Arthur Rimbaud used to test Reimagine's Generative AI tool
The original image we used to test Reimagine AI tool

AI generated versions

Model: Painting / Type: Vivid

This model and type are what Reimagine used by default. The model is likely based on what it detected, and the type could be one of “subtle,” “vivid,” or “wild.” Here is what Freepik produced immediately after uploading, without any button interaction.

AI generation of the Image of Arthur Rimbaud using Freepik AI tool
Image of Arthur Rimbaud using Freepik AI tool using “Painting” model.

As you can see, the result is quite decent, though not the same. It clearly failed with the eyes, both in color and expression. I tried using “subtle” and “wild” settings, but the outcome was the same. There’s a reason for this, which I will explain later in this article.

Model: Photo / Type: Wild

AI generation (photo model) of Arthur Rimbaud's painting using Freepik's generative AI
Arthur Rimbaud’s photo model from Freepik’s generative AI

Contrary to what I initially thought, the photo model is actually quite impressive. It still struggles with the eyes, although it’s more believable than the painting. The results are varied, and in fact, I expected more variety since I used the “wild” mode. However, the first and third images in the top row could easily pass the test. Keep in mind that the eye color of the original image is unnatural. I expected a more strict adherence to the image prompt for a picture, but for a photo, a natural eye color is acceptable. Had it taken the same color as in the painting, it would have looked like an alien.

Best AI model

Photo Shoot / Type: Subtle

After testing many models and modes, I have found that the best generative model for this particular AI image generation was “Photo Shoot” using “Subtle” mode. It’s closer to the original than all other tests, and it even captured green eyes. Although no model/mode combination could replicate the expression, the overall result is impressively good.

Image of Rimbaud using generative AI tools
Are we looking at Rimbaud’s real face? AI says yes.

Before anything: A word on UI and UX

Before conducting more tests, let’s examine how Freepik’s Reimagine generative AI tool operates, as it clarifies how the images are generated, what results to expect, and provides some insights on the overall user experience.

Reimagine’s UI (User Interface)

Freepik’s Reimagine UI is simple and to the point. Anyone familiar with Freepik knows that the design quality is consistently high, and its artificial intelligence section maintains this standard. It adheres to the design guidelines from the last Freepik rebranding, which is appreciated since the website had become complex with added features. Maintaining a known style guide rather than adopting a new one ensures a consistent user experience.

The dialogs are clear and direct, and the placement of each element is well thought out. For example, on the opening screen, the options are present but unobtrusive, allowing space for the main task. This is a smart implementation of Hicks’ Law, which enhances usability by reducing cognitive load.

All controls and options are intuitive and their functions are obvious. They also work immediately, eliminating the need to click a button to regenerate images. This convenience saves a click but also highlights a limitation: there is no history. If you prefer a previous version and accidentally change the settings, it is impossible to retrieve that version. Moreover, these changes disable the back button behavior, which is an important aspect of UI design.

Despite this shortcoming, which I hope will be addressed soon, the user experience is notably good, as I already have mentioned.

Reimagine’s User Experience

Quite honestly, this tool ranks among the best in terms of user experience for generative image tools. The interface is extremely simple to use; users understand what to do instinctively: there’s just a large space instructing you to upload or drop an image—nothing else is required, making the process crystal clear.

After you upload an image, some controls appear. These controls are present from the beginning, but the user interface is so well-designed that they can be easily overlooked. Missing them initially isn’t detrimental—it’s secondary to the main action.

When the aforementioned controls and options become visible, you will see various settings, including the prompt, the style, and the type or mode (which Freepik’s Reimagine refers to as “imagination”).

As for the image manipulation process, this is where I find the bigger issues: the model takes the uploaded image as reference, then extracts a text prompt (which is surprisingly accurate). The text prompt includes clusters of keywords, and the only way to modify that text prompt is by clicking on those keywords. This means that it’s not possible to add or modify anything that is not in the image already. And it also means that if the application does a bad detection, it cannot be modified either.

However, the Reimagine generative AI tool is part of a bigger set of AI tools (which I will explore in a future post), which include AI image generation, so users can use that option instead.

The overall user experience is really good, and the results are very satisfactory. Image size is really big (bigger than the original in some cases, with lossless size upscaling) and the image quality is second to none.

Second generative image test: Photo to Photo

For this test, I downloaded a premium photo from Freepik. In this case, a group of happy people. I know this is terribly difficult for AI generation, and the fact that Reimagine analyzes the image and creates a text prompt instead of working over the original image made me think I’d find problems in this test.

Original image to test AI tool capabilities

As you can see, there are two girls and two guys. Reimagine identified this photo as follows:

Reimagine's prompt on previous photo

The results quickly show the issues with this “image recognition to text prompt” approach. Just like in the painting by Rimbaud, it altered biometric features such as hair and eye color. It also failed to recognize one of the women in the prompt, although it succeeded in the generation. Despite using the “subtle” mode, the differences between the persons were noticeable.

As expected, problems with hands and ghost limbs arose. This is a well-known issue with AI, and I anticipated it when choosing this photo. However, the problems were not as severe as those seen with Stable Diffusion or Adobe’s generative AI.

Ai image generation using Reimagine
Reimagine’s Ai generation of the previous example image

For this generation, I used the “Photo” model set to “Subtle”. It’s worth noting that most other models provided worse results than this one, so it seems users need to find the exact combination depending on the uploaded image, which is to be expected from a generative “text prompt” image.

Meme and Text recognition using AI generative models

In this last test, I used a known meme with text to test Reimagine’s ability to recognize the content and preserve the image quality. My hypothesis was that since this meme is widely known, the text-to-image generation process would likely have problems with the facial features, but would be able to recognize the text correctly.

The original image I used looks like this:

original image used to test text-to-image AI prompt
Original image used to test text-to-image AI prompt

This is how Reimagine recognized the image:

Reimagine's text prompt interpretation of original image

Just by reading the prompt, I didn’t expect much adherence to the original. But I was curious about the text recognition. It also failed. Maybe not as much as most generative AI tools on the market, but still a failure. The following image is the best result I could achieve. Remember there are 25 modelsactually 23, but one is a “no model” option and another is an “all models,” which outputs image generations from all 23 styles at random—so there are 75 options to test. The following is the best one; for this, I used the AI style “Photo” and type “Subtle.”

generative AI image using "text to photo" prompt
Meme and text generation output

It’s easy to see that image generation failed because it depends on text rather than the actual image. Text recognition is suboptimal, but I can live with it; it’s not like any other tool gets it right, and I’m sure it will improve.

Final Test Insights

As is common at Dorve UX, we conclude each test with the insights obtained through research. Here are my thoughts on what could be considered “benchmark testing”—I didn’t conduct a formal benchmark, but given my familiarity with most tools and the extensive research I’ve documented in the previous article on Generative Image Tools, you might say I effectively benchmarked them, even though you won’t find results for the other tools on this page.

Before delving into the insights, I’ll share my expectations from a user experience perspective, which I believe is necessary to understand the final conclusions.

I discovered Reimagine through an email they sent, promising to enhance any image uploaded to their platform. Upon visiting the website, I noticed its simple, excellent interface, which I would rate 10/10. I couldn’t think of anything to REALLY improve; changes could be made, but that would be subjective—the interface is perfect as it is.

However, the app’s purpose is to improve uploaded images, and here we must be discerning. Did it improve images? It’s debatable. The output is undeniably high quality, and for certain images, it works exceptionally well.

Yet, if the app alters the main features of the image, making it literally impossible to retain facial features, then I think the app doesn’t function as expected and is more of a waste of time than anything else.

The issue lies in the underlying model they use. It’s not “image to image,” but “image to text” and then “text to image,” (see Bibliography for more details) which doesn’t make much sense for an image improvement AI tool. As I mentioned earlier, the tool “reads” the uploaded image and then performs a text interpretation, maintaining some reference points from the original image. I understand the reasoning: the image will always be high quality because it’s created from scratch from trained models. However, it fails on faces precisely because of this approach.

There are several ways to enhance the user experience of this tool. Some are simpler than others. The simplest “develop in one hour” solution would be a notice informing users that if facial features are important, then the tool is not recommended. Testing 75 combinations of models and modes is pointless if it will never work.

There are more complex solutions that could improve the experience without changing the underlying model, as well as approaches that involve changing the model itself. However, I won’t delve into these options since I lack information on the technology, and they are beyond the scope of this article.

In summary, Reimagine is excellent for certain purposes but very inadequate for others, leaving no middle ground. However, I know Freepik is continually improving its products, so I’m confident this article will become outdated and inaccurate in the near future.

Very Important Note!

Generative AI tools: Is Freepik new Reimagine any good? 1

Following a review by our CEO, Fabio Devin, it has come to my attention that Reimagine’s primary function is to create variations on images, rather than simply enhance them. Additionally, the suggestion I made regarding a notification for users concerned about facial recognition preservation has already been implemented by the time I published the article.

Therefore, I must correct my previous assessment and adjust my rating of the application to a 9/10. (-1 for Griffindor Freepik for those additional limbs!). I’ll leave my erroneous analysis for the sake of honesty.

Bibliography

We can improve your business!

Let us help you with the best solutions for your business.

It only takes one step, you're one click away from getting guaranteed results!

I want to improve my business NOW!