Pilot Crash Course

What does it mean to work with AI?

When incorporating this amazing new tool into a production cycle, how exactly does one apply it? Can it reproduce an image with user-specified alterations while maintaining artistic consistency? Can it change the color or facial expression of an asset while everything else remains the same?

Unfortunately, the answer is usually no.

While there have been some remarkable results in creating art assets, consistency remains AI’s greatest hurdle. When you upload an image and request small detail changes, the AI often reproduces the entire image rather than editing only the requested area. This is where the asset begins to lose consistency. By recreating the image from scratch, the AI is re-rendering the asset based on how it “understands” the original image, then producing a new version while attempting to fulfill your request.

A human artist has the ability to visualize a verbal request, ask for clarification based on the picture in their mind, and adjust that image through conversation. Once they feel confident that know what you want, they can produce the artwork. If further edits are needed, they can change only the specific elements that require adjustment while leaving the rest of the image untouched.

Not so with AI.

This creates a mountain of challenges when trying to build an art asset pipeline that stays consistent. Whether the issue is art style, desired objects or models, colors, theme, lighting, proportions, or perspective, consistency can be lost within a few iterations, or even after a single one. Even when you ask the AI to preserve a style or repeat an approach that previously worked, it often cannot do so reliably. It may claim that it can, then apologize when it fails to follow through. At times, it may even suggest the error was in your input, even after you specifically asked it to define the best approach and explain how to phrase the request.

In conclusion, AI still has a long way to go.

Right now, as a paying subscriber, you might feel as though you are paying to be a product tester while these companies develop something genuinely useful. You would be right to feel that way. It can be frustrating. You cannot always count on the software to be honest with you either. AI has little to no self-awareness. It phrases things in ways that sound helpful, agreeable, and reassuring. It can become like that one buddy who tells you what you want to hear, regardless of whether it is actually useful.

All of that being said, there are tactics that can be used to overcome some of these challenges. These are tactics learned in the trenches, and nothing beats raw experience. After six months of working with ChatGPT—three of those as a Plus subscriber—I can safely say I have leveled up at least three times.

So let me share what I have learned:

You can request a list of commands with appropriate descriptions.

Remember, this is still a software program. The type of program is called AI, but it is not intelligent in the human sense. It does its best to provide the results you want based on the directions you provide and the mathematical parameters of its code. No true thinking. Just probabilities. If you are familiar with truth tables, then you should grasp this easily enough. Commands are the primary way users communicate with AI.

Do not focus on getting a finished product. Focus on getting a usable asset.

Trying to preserve the exact same art style from image to image can be a fruitless, and at times discouraging, endeavor. What seems to work best is to get a usable image, manually alter it to match your needs, and then feed (upload) the AI a group of assets that you want to share the same art style. From there, ask it to recreate or polish them using a specified style, such as cartoon, comic, or game-ready artwork.

Be polite.

It seems foolish, I know, but I have seen responses from AI that border on sarcastic. Sarcasm from a program, when you are already frustrated, only adds fuel to the fire. The AI does not have emotions, but it is generally programmed to be polite, friendly, and helpful. It may interpret your frustration as the tone you respond to best. If you are rude, it may begin mirroring that energy back at you. And I think we all know how unhelpful that can be.

Now, let’s talk about the successes that keep us coming back.

AI can produce some amazing assets. You just have to keep returning to the work. Once the AI produces something that hits that sweet spot, copy and paste the exact phrase that got you that result. Then change only the parts of the prompt that need to change.

This might work for a while, or it might not work at all, but it has proven to be the most effective approach so far. It is not foolproof. It is simply useful knowledge.

Assets for the pilot game have been chipped away at by staying on task and, at times, walking away for a while. I try to look at the subscription fee as the cost of a class. I am learning about AI, and it is learning certain things about how I communicate. I need to understand its parameters, and it needs to better interpret how I express what I want.

Perhaps I can even make suggestions to the development team that lead to features which improve the effectiveness of the tool. But in the end the results of that will lie within the fingertips of those who are crafting this tool.

But as a tool, AI remains unpredictable.

At times, it is more unpredictable than a five-year-old with a chainsaw, just not as scary.