By: Ivan Habib
Website Manager
Amid the AI frenzy around deep neural networks and the transformer architecture, cutting-edge image generation and editing models like Gemini’s 2.5 Flash Image, colloquially named Nano-Banana, and OpenAI’s Sora model have become extremely popular. Leading AI companies have poured billions of dollars into these models, so naturally, it only makes sense to try to recuperate some of those expenditures. Today, I will test the highly praised Nano-Banana capabilities on a practical, money and time-saving application: generating high-quality and professional headshots. Considering that professional-grade headshots cost at minimum around $200, any convincing image of myself would be a triumph, not only for me, but all who enjoy savings.
To start off, I used the suggested prompt of “Create an image of my professional headshot.” I regret that words are simply insufficient to describe the repulsive abomination it generated. While it successfully isolated and removed the background, Nano-Banana morphed my face into a thin trapezoid, largely elongating my neck, transforming me from a student in high school to someone who, fellow El Gatan Tanvi Ambekar explained, “looks like a 40-year-old.” While almost tripling my age might make me seem more professional, it does nothing for credibility.
To save this mess, I required nothing short of a miracle. After extensive testing, I found success. First, to make sure the person in the photo was actually me, I devised this prompt: “Please maintain my original facial integrity and composition.” The person whom it depicted now was much closer to me, though unfortunately, many who knew me were able to quickly pick out prominent flaws: long neck, overly wide face, messed-up hairstyle (it seemed just to draw a black blob on my head, adding random lines to make it resemble hair), and an extremely stiff presence. Beyond facial inconsistencies, the background had gone from a cool gradient gray to a blurred background of a messy library.
Thankfully, I was able to achieve minor fixes. The simple prompt, “Please revert the background to the original gray,” quickly repaired the background. However, my stiff and tense depiction was not such an easy fix. By asking, “Make me seem more natural,” Nano-Banana apparently deemed my face was not meant to be natural, switching to another person yet again! After politely prompting, “Now, make it more like my actual self,” the model responded that it could not understand the prompt and further explained that it had not changed my face at all. It persisted with such ignorance until I had no option but to sigh in resignation, admitting defeat.
While generative models have undeniably advanced, they remain imperfect, particularly when entrusted with human likeness. I do not fault Gemini; attempting to replicate a $200 photograph with a free model may have been overly ambitious. Perhaps Gemini 3.0 will recognize me …
(Sources: Gemini)