How To Get Good Quality Complex Images - Part 1
By Arcturus
A complex image is an image in which many elements contribute to the harmonious effect of the whole, that is, the image is enjoyable only if all the elements contained in it are like the voices of a choir singing in harmony. Getting good quality complex images can be a challenging task, as we will see in this tutorial. Sometimes complex images are born from ideas that arise in the minds of creators, without knowing the exact details of the image they want to see, but having in mind the particular effect and atmosphere they want to achieve. AI models may surprise them with particularly creative solutions, or may deceive them with poor results. In any case, users know very well whether the desired effect has been achieved or not. For example, let's suppose I want to get the image of a "Dreamland City", an otherworldly place where surprising shapes of architectural structures and strange plants and gardens blend into a photorealistic image, which however clearly shows a magical, almost paradisiacal environment, certainly not belonging to our world. The first process I'll use is text-to-image, since I don't have any reference image yet, but just an idea in my mind. In the second part of this tutorial we'll see how to improve and enhance some of the images obtained with this process. So I prepare a good prompt in these terms: "Create a cinematic, breathtaking panoramic view of a hyper realistic otherworldly city made up of precious marble buildings of various colors and bizarre curvy forms with golden decors and many windows and arches, joint together side by side and one over the other. Between them there are amazing gardens, full of wonderful alien plants and very strange trees with stunning forms. Dark velvet sky with a Milky Way (galaxy) and two colored planets. Paradise dreamworld, phantasmagoric atmosphere, amazing digital art with lots of details, sharp focus, vibrant color palette, warm cinematic light, HD, masterpiece, award winning photo, extremely detailed octane render." BudgetPixel gives us the ability of using all the most important text-to-image models currently available, so let's examine the results I got with a dozen of them, always using the same prompt as above. But first I want to show you three final images that represent well the effect I want to achieve. Let's now look at the results of the various models with a brief comment under each of them. DALL-E 3 - Standard quality - 80 credits. Not what I want 👎 Flux 2 Pro - 25 credits. Not photorealistic, low contrast, some creativity but not well developed 👎 Flux Kontext Max - 80 credits. Some of the shapes are interesting, but overall the effect is more like an illustration from a fairy tale than what my prompt called for. GPT-Image-1.5 - 60 credits. Not so bad, could be selected for the second phase of improvement and enhancement 👍🏼 HiDream I1 Full - 30 credits. Too plastic, cartoon-like. Hunyuan Image 3 - 60 credits. The idea has been developed to some extent and could perhaps be selected for improvement. Ideogram v3 Quality - 90 credits. A perfect example of how even the most expensive models can give very poor results 👎 Imagen 4 Ultra - 60 credits. Absolutely disturbing 👎 ImagineArt 1.0 - 10 credits. Cheap model, cheap result. Lucid Origin Standard - 18 credits. This model has some settings that need to be tested and, being cheap, it can offer some interesting solutions, as you can see in these three images. The third one is my favorite. Lucid Origin Ultra - 85 credits. Expensive, and photorealism here becomes a kind of boring realism. Nano Banana 2.5 - 40 credits. Candy-like, I'd say! Qwen-Image - 15 credits. Too simple and cartoon-like, even if the idea has been developed. SeeDream 4.5 - 45 credits. This could be selected for enhancement 👍🏼 SeeDream 3 - 30 credits. This image has potential and could be improved. SeeDream 4 - 33 credits. One of the best images of the series, definitely to be selected for improvement 👍🏼 Wan 2.6 - 30 credits. The two images above are not well focused, but can be improved a lot 👍🏼 Z-Image-Turbo - 15 credits. Another cheap model that creates images which can be improved to a good quality level 👍🏼 So, that's all for today. Now we have 4 or 5 complex images that we can use in the image-to image process to achieve the desired quality level.