Until not long ago, converting an image like the one on the left into something like the right was not entirely simple. You could do it, yes, but for this you needed advanced knowledge of image editing programs such as Photoshop and, of course, time. That is increasingly true, and using the generative AI to transforming images is already a clear example that this technology can make what was very difficult before.
Hello, Qwen-Image-Editor. This is the name of the generative the AI model Posted by Alibaba And that is part of his increasingly reputable family of models Qwen. We are facing a specifically trained model to edit photos and images extremely simple: it is enough to ask for a prompt.
Using this AI is much easier than using Photoshop in a traditional way, and in many cases the results are exceptional.
You send, the AI obeys creating. The new Qwen-Image-Edit is available to try directly On the website of this model. When we go to that page we will see how the “Image Edition” mode is already preloaded, and it will be enough We can also work on that image successively.
An increasingly capable model. This model is a kind of ‘Fork’ of Qwen-Image, the founding model of 20,000 million parameters (20b) that Alibaba published at the beginning of the month. This model focused on things how to render text in the images, but here the edition goes further. The model is available in Qwen Chatbut also in Hugging Face, Modelscope, GITHUB and through the API of Alibaba Cloud (The cost is $ 0.045 per image, but the first 100 are free).
Source: Alibaba Cloud.
In the cloud and at local, on your PC. Although it is very comfortable to use it from the Qwen Chat interface – which allows you to generate up to 8 free images (entry/exit free every 12 hours -, it is also possible to download it and use it at a home on a PC or laptop with sufficient graphic memory (the model weighs 54 GB).
The expert Simon Willinson has done it in his MacBook Pro M2 of 64 GB of unified memory and verified how with 10 inference steps (by default he uses 50, but that takes much longer) It took 25 minutes in generating a modification of an image. In Qwen Chat, which takes advantage of much more computing resources, you just have to wait a few seconds for each image.
Understanding the image before editing it. One of the advantages of Qwen-Image-Editor is that it makes use of the so-called double coding mechanism. This system consists in analyzing the input images with the Visual Recognition model QWEN2.5-VL to “understand” the image and know what is in it and then use a variable autocodifier (VAE) to add the details of lock (the edition). This approach manages to maintain fidelity to the original modifying only what the user asks.
Plate with hair, dish without hairs. The difference is subtle, but very relevant, especially because you see the rest of the image absolutely intact.
Junyang Lin, one of the researchers involved in the development of this model, explained That allows how “can eliminate a hair from a plate” leaving the challenge of the un modified image, something that according to his words “is a very delicate image modification.”
The Prompt indicated to the model that changed blue only the color of the letter “N”. So the model analyzing the original image, recognizes that letter, and changes the color without modifying the rest.
Semantic edition. Another of the great characteristics of this model is its ability to make a semantic edition of the image. This is: what changes is the structure or meaning of an image. For example, we can apply the famous image effect with Studio Ghibli or another that transforms a photo In a scene created with lay peoplebut preserving the identity of the objects or people of the original image.
Transforming any image into another that seems like a scene of some Lego movie is also very simple with this model. And again, the original content maintains its essence.
Transforming reality. Like other similar generative models, evolution remains remarkable. We already saw how Google’s magic editor has long been raised, but Qwen-Image-Editor goes one step further to be able to maintain that fidelity of the appearance in the original image.
For example, adding graffiti in English or Chinese – Diomas that this model manages perfectly – while the rest of the image remains identical to the original. The quality and consistency of the texts is extraordinary. They even adapt to clothes, even if you have wrinkles, so that the realism effect is greater, as is the case (although it does not notice so much) with the image of the image of Mark Zuckerberg.
An example of a prodigious (and disturbing) trend. Here we are talking only about a model that edits images and transforms them, something we had already seen in the past but that with Qwen-Image-Editor goes further. The lesson is actually deeper: all these models manage to make us glimpse a future in which to manage to handle complex applications (such as Photoshop and the like in this case) will no longer be (so) necessary.
Learn Linux? I? Except in specific cases, many of the things we did with these applications can be done without having any idea of graphic postprocessing: it will be enough to “talk” with an AI to ask for what we want. And what is true for Photoshop will be equally for endless tools and systems: suddenly squeeze to the maximum Excel or the Linux console – for example – it will no longer make us necessary to study them and use them for hours (even if that always helps). AI will have already done it.
GIPHY App Key not set. Please check settings