Watercolor

I have always liked watercolor painting. I greatly admire those artists who – especially in the nineteenth century – mastered the technique of skillfully balancing color intensity and the application of glazes, achieving stunning results with their confident brushstrokes. Unfortunately, I must confess that I have never quite lived up to those masters, even though I do manage to produce something decent now and then. Unlike other painting techniques, such as oil or acrylic, watercolor is unforgiving, and mistakes cannot be corrected. One of the most famous watercolorists, whose works I have always admired, was Ettore Roesler Franz (1845-1907): he captured many scenes of 19th-century Rome "on the spot," before they were lost forever in the mists of the past. Here is an example of his skill as a painter of cityscapes: around 140 of his watercolors can be seen today in a permanent exhibition at the Museum of Rome, not far from Piazza Navona. For these reasons, I wanted to experiment with AI to see which models were best suited to capturing not only the effects but the very spirit of watercolor painting. I began my research by deriving a prompt from an old photograph to use for text-to-image generation, ensuring the model would not be influenced by a pre-existing reference image. Here is the prompt I used: "Watercolor illustration on white textured paper; edges slightly faded. An elderly man with white hair and a beard, wearing a brown leather jacket with a shearling collar, stands motionless at the edge of an old stone-and-wood canal lock. He gazes down into the still water. Canal water: opaque depths, distorted and faded reflections of the wooden gates, translucent patches of bluish-grey and ochre, faint ripples, no specular highlights, and color gradients blending without hard edges. Style: traditional watercolor; fine, light black ink lines for outlines only; soft shading; visible paper texture with slight grain; desaturated palette of earth tones and greys; flat, twilight atmosphere; no digital effects. Composition viewed slightly from above; low horizon line; pale, empty sky." Below, I will briefly comment only on the most interesting results obtained with the various models tested. Above, SeeDream 5: the watercolor effect is rather good, but the image is poor and the textured paper effect is not sufficiently emphasized. GPT Image 2.5 (quality: High) is much better; the paper grain is clearly visible, and the landscape has been completed in a way that is consistent with the subject's atmosphere. Here is the result for GPT Image 2.5, Medium quality (cheaper), without changing anything in the prompt. In my opinion, we are at the same level as the previous image. This is the result with Qwen 3.0 Pro: the edge of the paper is clearly visible and the watercolor effect is satisfactory, but the image is – once again – poor in quality. With Grok Imagine 2.0 (Medium quality), we are more or less on the same level, although the image turns out more complete. This is the same model, Grok Imagine 2.0 (Low quality): the paper grain is reproduced very faithfully. Other models, such as Muse, Nano Banana 2 or Ideogram, gave very poor results, which are not worth showing here. Perhaps the best watercolor effects were obtained wit Gpt-Image-2 (Medium quality). Here they are: In conclusion, if you love watercolors, you can achieve excellent results with GPT-Image-2 (Low or Medium quality) by using a suitable prompt and, optionally, an old black-and-white photo as a reference image. Here are some examples of the effects that can be achieved. If you have a good color printer, some of them could be worth printing!