Riverside Dreamscape
Tuesday. A beautiful start to the day, and no work for me. Dash has RE and Biology papers today, so I dropped him off at school at the usual time. Back at home I got a few jobs out of the way, and went for a short walk on the lanes; then, in the afternoon, settled in for a bit of photo-editing. I think I’ve solved my ‘death-of-Photoleap’ problem: the app itself is still going, but they removed the old AI Selfie mode which was great for producing interesting abstracts when used on images that did not contain a face. A real shame, but to be fair, mine was a very niche use-case for the product. So, I decided to experiment with some of the newer image generators to see if I could replicate the process — keeping the composition and structure of the original photograph, but rolling the dice on colours, textures and details based on a theme. Control mixed with a dash of serendipity. I used the following prompt:
“Create a surreal abstract image, dali-esque in style, based on the uploaded image, keeping the underlying structure and composition but rendering the elements otherworldly and surreal ; feel free to be quirky and imaginative, and aim for a more painterly or artistic style, no need to stay photorealistic, and it's fine to completely change the colour palette - aim for striking and vibrant.”
Both images start from a very ordinary documentary photograph: a straight, architectural view up the ramp to The Riverside, with its repeating wooden cladding, railings, and the delivery cyclist providing a small focal point. What makes the experiment interesting is how differently ChatGPT Image 2 and Nanobanana 2 interpret exactly the same structure and composition once I asked them for something Dalí‑esque and surreal, but still grounded in the original frame.
The offering from ChatGPT Image 2 (above) feels closer to a classical oil painting, almost like an alternate‑universe version of the original photo. The perspective lines and basic geometry stay faithful: the railings, the reflections, and the vanishing point all read very much like the source image, but they begin to sag and liquefy in ways that directly nod to Dalí’s melting clocks. The wooden cladding on the right is transformed into thick, flowing ribbons of timber that behave more like wax than wood, hiding eyes and watches in their folds. The palette shifts to rich golds, teal skies, and warm shadows, giving the scene a late‑afternoon glow that feels contemplative and slightly ominous. It’s as if time has started to run in slow motion along the riverside, and the everyday architecture is quietly dissolving.
Nanobanana 2 takes a much more exuberant approach: I’m not entirely sure what it has been smoking! It keeps the same compositional spine — the ramp, the railings, the building angles, the cyclist heading into the distance — but everything else is dialled up into neon psychedelia. The sky becomes a swirling vortex, the path turns into a ribbon of galactic colour, and the building itself mutates into a dripping, animated façade. The vertical slats on the right no longer read as wood at all; instead they become crystalline, jewel‑like shards, each one inlaid with tiny surreal details: eyes, skeletons, pocket watches, and flashes of colour that feel closer to a graphic novel or festival poster than a traditional painting. Where ChatGPT Image 2 suggests a dream slowly bending reality, Nanobanana 2 looks like the same scene after someone has turned the psychedelic volume knob all the way to the right.
For me, that contrast is what makes the pair so compelling. Both models respect the underlying composition of the original photograph, but they project very different personalities onto it. ChatGPT Image 2 leans into painterly brushwork, warm light, and a restrained set of surreal motifs to create something that could almost hang alongside mid‑20th‑century surrealist canvases. Nanobanana 2 embraces maximalism: high‑saturation colour, dense visual symbolism, and a more illustrative, contemporary feel. Rather than proving that one approach is “better,” the two images together highlight just how much expressive range you can get by feeding the same source and prompt into different AI systems. Even in the world of generative tools, style is still a matter of interpretation.
