Hello! I am not questioning the advantages of the Stable Diffusion neural network. However!
I charge by the hour for my work with neural networks, as I have decided. I tried to calculate the estimated cost of the project as a whole. In rare cases, it is also convenient. But neural networks are often still a black box. And you need to make a lot of attempts, look for suitable prompts and styles.
Once, when I received an order to make a poster for a TV series that would later be released, based on a novel, I thought, okay, one poster, I’ll spend about a day on it, the price is such and such, we signed a contract, taking into account corrections, etc. In the end, I worked on the assignment for three days, but calculated the payment as for one day. I decided to calculate the cost of my hour of work and offer clients exactly those terms.
I am currently working on an order on the theme of fairy tale characters in modern office settings. The customer allowed me to publish some of the generated images.
This time, the order is quite complex, with many corrections and additions. I work in the Midjourney V6 neural network, and the Vary Region function helps me edit the image. For example, at the beginning, I got Little Red Riding Hood in the server room on a red background. But it was decided to change the background to blue for contrast. Vary Region handled this task with ease. I selected the character with the lasso tool and wrote a new prompt for the background.
An important task from my client was to create a fairy-tale atmosphere, a digital image (not a photo), bordering on digital art and photorealism. And, of course, the speed of execution. The order is for months, and we spend from several days to a week or more on each image. There will be a lot of images at the end.
I decided to test the Stable Diffusion (hereinafter SD) neural network for this task. I hardly ever use this neural network. And, of course, I was convinced that speed and quality are not SD’s strong suits. Yes, you can train the LORA for weeks, develop a character, or use ready-made LORAs, but even if you download a cool LORA, you still won’t know all the secrets of the prompts that the LORA creator wrote. In the end, I realized that with SD, it would take a lot of time to achieve the necessary image aesthetics. In my opinion, Midjourney ranks first among neural networks in terms of quality and speed.
These are the images I got in the SD neural network with a good quality lora, but without finishing the image, as I don’t have the time or desire to mess around with SD settings. We live in an age of fast content, advertising, and news stories. I don’t see the point in bothering with extra work. This is just my opinion. If people like it, WHY NOT? And I can draw for pleasure as much as I want.
As we can see, the result is more photorealistic, but there are just as many flaws, if not more, than in Midjourney. There is still a lot of work to be done. I left SD alone. I continued with Midjourney, especially since I have been working with this neural network for a year and a half and have mastered it quite well.
Yes, the Magnific Ai neural network copes well with problems with fingers, eyes, and skin—it has its own settings and is easy to use. The output quality is surprisingly good! Below is an example. I ran an image generated in Midjourney through Magnific Ai.
