I have seen some people online remastering old games using AI, and I have always enjoyed reimagining classic titles myself (I even did a project for Mega Man a long time ago). Recently, I felt inspired to try this with Final Fantasy VI. Because the original game uses 2D sprites, it requires a fair amount of imagination to visualize the world. I wanted to see what the results would look like if I illustrated various scenes in a semi-anime style.

It took me an entire day to complete, but I gained a lot of experience and had a great time doing it.

Images from the Original Source

I searched for original gameplay screenshots online to use as references. Because I did not note the specific sources when I saved them, I apologize for not being able to provide proper credit to the original uploaders.

I tried to choose iconic scenes that players would likely remember.

The Generated Images

The main process involved creating images using Google Nano Banana via Gemini. However, it is rarely possible to get the desired result in a single attempt. It usually requires multiple prompts or creating several variations and combining them afterward in Photoshop.

And here are the result:

Since I’m already used AI, I decided to try generating a video as well. I then processed the generated image into Kling AI (though I actually feel Veo 3 produces better results.

Character Sheets

I used a method of creating a character sheet beforehand. When generating a full scene, I uploaded the character sheet as a reference to ensure the character’s appearance remained consistent.

Sabin: His design does not quite match the original in some areas, specifically the colors of his shirt and pants. I tried to generate armor that matched the original sprite but couldn’t get it right, so I eventually gave up on that detail. For his pants, I felt the yellow color looked better, so I decided to keep it.

The part below wasn’t used. 

Terra: I was unable to use the prompts for Terra’s Esper form directly because Gemini considered the design too revealing, which caused generation errors. It actually required a significant amount of manual editing to achieve the final look.

Celes: In the original game, Celes wears a green leotard. I tried creating a version without anything underneath, but it looked too suggestive for the game’s tone. However, adding layers underneath looked weird too, so I changed her outfit to a skirt instead. I think this suits her better. I also considered white shoes (which are closer to the original) for the middle image, but by then it was too late to change the whole set. I think the blue ones look better anyway.

Overall Impressions of Generative AI

Generative AI is currently very capable; it can create professional-level images. However, there are several notable limitations:

1. Lack of Precise Control: Certain styles or specific details are difficult to achieve. Even after repeated prompts, the AI often fails to produce the exact intended result, which can lead to exhausting your available credits.
2. Repetitive Aesthetics: Character faces and body shapes often look nearly identical across different generations.
3. Lack of True Innovation: The AI struggles to create truly “out-of-the-box” or highly creative concepts.
4. Visual Distortions: Upon closer inspection, strange artifacts and distorted details are still common.

Furthermore, the more you instruct Nano Banana to modify an original image, the more the overall quality tends to degrade, except for the newly modified sections. To avoid this quality loss, you must either keep the original base and overwrite parts manually in Photoshop or generate an entirely new image from scratch.

While AI is useful for brainstorming or performing simple, repetitive tasks, it definitely cannot replace a real artist at this stage.

Conclusion

Overall, this was a very fun experiment. I will definitely try this again if I get the chance.