Tip 10: Progressive Optimization with Controlled Variables
Many people liken the process of AI generating images to alchemy, which I think is very fitting. Often, for some unknown reason, adding a mysterious formula in the prompt can result in vastly different images.
However, although the process may resemble alchemy, it does not mean that we need to resort to ancient methods and rely on luck. In my opinion, the best way to use Midjourney currently is through progressive optimization with controlled variables.
Taking the example of the equipment set image above, if I want it to generate equipment arranged piece by piece and have a unique effect on the sword, first we need to systematically examine each line of our prompt.
| prompt | check | |
|---|---|---|
| Type | blender 3d, game sheet | This seems to be a bit of a problem. The style I was hoping for in Hearthstone is not this kind of delicate 3D effect, but rather a style more similar to Starcraft's, while Hearthstone has that clay render style. |
| Main Body | lightsabers, paladin armor, | There may also be a problem here, perhaps not writing 'different types of lightsabers' resulted in only one type of lightsaber being generated. |
| Background | null | This should not be a problem. |
| Composition | null | This should not be a problem. |
| Lens | null | This should not be a problem. |
| Style | style of Hearthstone | I feel that the generated image lacks the greasy feeling of medieval armor. But if we change the style to clay, would it be better? This area may also need some optimization. |
| Parameters | null | This should not be a problem. |
From this table, we can see that we have identified three issues. The first issue is related to the clay style and the last issue. My suggestion at this point is:
Only modify one variable at a time while keeping the others constant. After modifying one variable and confirming that it meets expectations, move on to the next one. In my experience, it's best to prioritize modifying the main content because it may affect other variables. If you want to ensure consistency in your modifications rather than just coincidentally meeting requirements, you can generate another prompt using the same method. However, since generating each prompt incurs costs, this is an optional step.
Okay, following these principles let's modify our prompt and see what gets generated. After adding 'different types of', we now have more lightsabers - problem solved.
Then we modified Blender 3D to change it to clay render. After I finished the modification, I found that the texture of the generated results (the left four images) seemed a bit lacking. So I tried preserving the oily effect in 3D Blender (the middle four images), as well as using both clay render and oily effects together (the right four images).
I personally feel that the best effect is achieved when both are changed (the four on the right).
If you compare the images generated by the initial prompt, isn't the final image better? So this technique mainly wants to tell everyone that there is no need to copy answers when writing prompts, and you have the ability to write good prompts yourself.