[Some Images] Testing out the new chatgpt model

Hi there! I hope you’re doing great. If you’re into AI, especially AI image generation, you might have heard that OpenAI released a new image generation model last week. I had the chance to test their previous model when it came out, and I must admit, I wasn’t particularly impressed. It was quite restrictive and didn’t produce high-quality or usable images, at least for my standards.

However, let me tell you about this new model. I won’t bore you with the details, but I’ll give you a brief overview of what I liked and didn’t like about it.

Things I liked:

• High-resolution, sharp images: The model generates images from the start in high resolution, eliminating the need for upscaling. The default resolution is 1024x1536, which is impressive.

• Excellent prompt adherence: The model follows prompts quite well, although it may require a bit of prompting with asterisks.

• Affordable pricing: It’s free to use, although there are limitations on the number of images you can create with the free tier.

• User-friendly interface: The interface is intuitive and easy to use, making it accessible to users of all levels.

Things I didn’t like:

• Creation time: It takes a significant amount of time to generate images, ranging from 30 seconds to a minute. I’m not sure if I’ve become accustomed to faster image generation times, but it’s definitely a downside.

• Incorrect prompts: If you don’t provide the correct prompt, the model tends to apply a strong yellow filter to the image.

• Limited representation of women: The model tends to portray women with more realistic proportions, which may not be ideal for certain use cases.

• Iterative process: You often need to go back and forth with the prompt to achieve the desired look.

• No option to add lora's or custom models or styles. Very little customization.

Now, let’s take a look at some sample images generated by the new model.

'create an image of a muscular woman inside an office'

 Pretty good results for a simple prompt, as you can see, proportions are more much realistic, very nice details, good lighting, realistic skin. I like it very much. But let's make some changes: 'make her younger, blonde, professional blouse and mini skirt'.

 I don't know what in this this new prompt made her smaller, but at least it got the clothes right. 'make her a little older and more muscular'

 I have noticed that with requested edits, it tends to be too extreme. It works for us to make the muscles bigger, but it also made her way older and more masculine. Nothing wrong with that but that's not the look I'm looking for. Let's try one last one. 'Make her more feminine and younger'.

Well, it not only fixed her clothes but also made her skin smoother and even improved her hair! I’m thoroughly impressed with these remarkable results. I believe that with a more precise and detailed prompt, you can achieve even better outcomes from the initial image. Additionally, I’m impressed with the scene consistency. For the most part, it adhered to my instructions and refrained from altering what I didn’t specify. This is a significant advantage compared to my current model, which can be quite challenging to work with. Consequently, I often resort to creating a new image and hoping for the best. Oh, and by the way, the hands and fingers are almost perfect! Congratulations, OpenAI!

Let's see some other examples:

'Give me a muscular woman, as a teacher, long black hair, inside a classroom'

As you can see, it has this very noticeable and now very chatgpt identifiable yellow look. I try correcting it: 'make her younger, more vascular, make the image bluer, no yellow filter, give her a white shirt, she holds an apple'

Well, it certainly made the images bluer, but that’s what I mean when it tends to overcorrect. However, it’s excellent at poses, realistically holding objects, and most images tend to be ‘usable’, unlike my current model, which yields a mere 35-40% usable image rate.

A couple more examples:

'Give me of another huge muscular girl in a crowded house party at night, wearing a short dress, holding a red cup, dark room'

 'make it slightly cooler, not so yellow, but don't over do it, normal white balance, flash effect, make her more feminine'

 And one last one:

'A woman at a gym, dramatic lighting, make her bigger, more vascular, short gym shorts, top, long red hair in a ponytail, freckles, feminine face, young, normal white balance.'

 This is a very good image that I got from the first try. So, as you can see, with the right prompt, you can get very good, realistic images without needing to do much editing afterward. I will buy the paid plan to continue experimenting with this new model and report my findings next week 🫡. Have you tried this model? What are your favourite prompts?

You can find all the images attached to this post.