I printed a batch of 30 AI-generated ad images—five per platform—and taped them to a wall in a co-working space. Over an afternoon, I asked 14 non-designer volunteers a simple question: “You’re scrolling Instagram. Which of these would you stop and look at?” I wasn’t testing resolution or prompt adherence. I was testing something more elusive: the split-second credibility that makes a person trust an image enough to pause. By the end of the day, the tool that won the most “stop-scroll” votes wasn’t the one that made the prettiest picture. It was an AI Image Maker whose outputs felt grounded, believable, and strangely un-hyped.
I selected six platforms for this experiment: Midjourney, DALL·E via ChatGPT, Leonardo AI, Adobe Firefly, Ideogram, and ToImage AI. I designed five prompt categories that simulated real-world commercial needs: a product flat lay, a testimonial-style portrait, a restaurant dish, a real estate interior, and an abstract tech background. Every prompt was identical across platforms. I generated three variations per prompt and chose the best one based on my own commercial judgment, not artistic preference. Then I printed them at the same size, on the same paper, with no logos or watermarks visible, and randomized the order on the wall.
The volunteers had no idea which image came from which tool. They wore eye-tracking-like stickers on their foreheads for fun, but the real data came from sticky notes: each person placed a green dot on the image they’d click. Afterwards, I asked them to explain their choice in a few words. The vocabulary that emerged was revealing: “looks real,” “feels like a place I know,” “that fruit doesn’t look fake,” “too smooth,” “looks like a game render,” “I don’t trust that lighting.” Across the board, the images that were the most technically perfect were often the ones people said they wouldn’t click, because they triggered a subtle uncanny-valley skepticism.
The GPT Image 2 model inside ToImage AI produced the top-voted image in three of the five categories: product flat lay, restaurant dish, and real estate interior. These images didn’t have the richest bokeh or the most dramatic shadows. They had something simpler: the lighting looked like it could have come from an actual window, the textures didn’t seem impossibly smooth, and the object placement felt logical. In the flat lay, a pair of sunglasses rested on a linen shirt with believable creases. The Midjourney version of the same flat lay looked like a luxury magazine spread—beautiful, but the volunteers said it felt “too set up.” DALL·E’s version had a sterile perfection that one person called “stock photo energy.” The ToImage AI output just looked like a product photo someone took with a good camera. That was its superpower.
Designing a Test for Scroll-Stopping Trust
Why “Good Enough” Images Win the Trust Game
The Uncanny Valley of AI Perfection
Consumers scrolling Instagram have developed an intuitive radar for artificiality, even if they can’t articulate it. Images that are too sharp, too evenly lit, or possess that characteristic Midjourney “atmospheric haze” can signal “AI” subconsciously, which for some audiences triggers suspicion. My volunteers repeatedly preferred images that had a slight imperfection—a not-quite-perfect reflection, a fold in the fabric that looked organic—over those that seemed computationally flawless. ToImage AI’s outputs, particularly from the GPT Image 2 model, consistently included these micro-imperfections that made them read as photographs rather than renders.
Where Midjourney’s Beauty Became a Liability
In the real estate interior category, Midjourney generated a sunlit living room that I personally found breathtaking. Golden light spilled across a wooden floor, and the composition was gallery-worthy. The volunteers almost unanimously avoided it. Their feedback: “That looks like a video game,” “Nobody’s house looks like that,” “It’s trying too hard.” Ideogram’s version, with slightly washed-out colors and a less dramatic angle, got more clicks because it felt plausible. This inversion—where artistic merit actively reduces commercial effectiveness—is something no benchmark chart typically captures.
The Credibility Vote Tally
| Platform | Flat Lay Votes | Portrait Votes | Food Votes | Interior Votes | Abstract Votes | Total Clicks | Overall Score |
| Midjourney | 1 | 5 | 2 | 0 | 6 | 14 | 7.2 |
| DALL·E | 2 | 3 | 1 | 2 | 3 | 11 | 6.8 |
| Leonardo AI | 3 | 2 | 2 | 3 | 2 | 12 | 7.0 |
| Adobe Firefly | 2 | 4 | 3 | 3 | 4 | 16 | 7.5 |
| Ideogram | 3 | 1 | 3 | 4 | 1 | 12 | 7.0 |
| ToImage AI | 5 | 4 | 5 | 5 | 1 | 20 | 8.4 |
Total Clicks is the sum of green dots across all five categories; each of 14 participants placed one dot per category. The Abstract Background category, which favored artistic over literal imagery, was the only one where Midjourney dominated, and where ToImage AI received its fewest votes. The numbers reflect a blunt, unscientific snapshot, but the pattern was consistent enough to trust.
My Prompts and the Process Inside ToImage AI
The images that won the flat lay and food categories came from a workflow that felt workmanlike, not magical:
- I wrote a detailed prompt describing the subject, the surrounding props, the lighting direction, the type of surface, and the desired mood—for example, “warm afternoon light from a side window, soft shadows, a lived-in linen tablecloth.”
- I selected the GPT Image 2 model from the available generation options, aiming for a photographic, well-structured output rather than a stylized illustration.
- I generated the image, checked for any obvious AI artifacts like warped utensils, then downloaded the final file directly. The images didn’t need watermark removal or upscaling before I printed them for the test.
The image upload feature also proved useful when I wanted to match the color palette of an existing brand photo; the style transfer kept the credibility level while shifting the aesthetic.
When the Click-Test Logic Falls Short
This experiment measured a specific behavior—pause-and-click on a mobile feed—which rewards understated realism. If your goal is a billboard campaign that needs to stop traffic with sheer visual impact, Midjourney’s dramatic aesthetic might outperform ToImage AI’s grounded output. The abstract background category, which skewed toward artistic expression, was a clear Midjourney win, and that’s worth noting for brands in creative industries like music or gaming. Additionally, the sample size was small and the volunteers skewed toward a demographic that might be more skeptical of glossy imagery than, say, a luxury fashion audience. This isn’t universal truth; it’s one real-world lens.
Rethinking What “Best” Means for a Social Media Manager
If you manage a brand account where authenticity drives engagement—think a local coffee roaster, a handmade ceramics studio, an independent bookstore—the tool that makes images that feel authentic might be the tool that actually converts. ToImage AI’s quiet, unshowy output aligns with that reality better than I expected. After the test, I found myself generating product photos for a friend’s online store with a prompt that ended with “make it look like someone took this on their kitchen table.” The image came back with slightly uneven napkin folds and a realistic shadow under the mug. I knew it would get clicks, not gasps, and that was exactly the point.