The Marks I Would Have Removed

A portrait photographer’s unease at an image that looks real.

Café terrace © Philippe Triconnet, not AI-generated. The AI-generated portrait discussed below appears in the source article.

In an article published by HalfJourney Lab, a woman sits by a café window. Looking at that image, I cannot find anything that tells me she was generated. I know she was, because the article explains how the image was made. But knowing its origin does not help me see it.

What catches my attention is her skin. There are small marks I would probably remove when working on a portrait. Here, they contribute to making an invented face look believable. As a portrait photographer, I find that troubling.

In the article, the author takes the reader through several stages of image generation and transformation. Following those stages reveals how deliberately this photographic appearance was constructed.

The first image comes from Midjourney. The prompt describes a woman stirring her coffee while looking through a window, unaware of the camera. It specifies loose strands of hair, a grey sweater, a book, glasses and slightly imperfect framing. Even before the image appears, the circumstances of an apparently unposed photograph have been carefully arranged.

The author then feeds the Midjourney image into Qwen-Image-2.1, an AI image-editing tool, with instructions to make it look more photographic. The instructions ask for visible pores, fine skin texture and subtle freckles “only where they belong”, along with natural daylight or flash colour balance and visible 35mm film grain. They also exclude acne, spots, rashes and added redness. The desired skin is carefully specified, down to which irregularities are welcome.

I cannot trace each mark on the finished face to a particular instruction. But the texture that makes me believe in the image was explicitly requested.

The next step introduces photographs as visual references. According to the author, their influence is smaller than expected.

The larger change comes with a LoRA, an adaptation trained on a collection of analog photographs to steer the model towards their appearance. The author describes more texture and less perfection, at the cost of losing some of the original image’s distinctive features.

I cannot judge the whole process from a few selected examples. I have not tested the setup. But I can describe my reaction to this particular result: viewed on screen, I would accept it as a photograph. I cannot point to a decisive visual clue that would tell me otherwise.

In When Your Work Becomes Training Data, I wrote about my discomfort with photographs becoming training material without their authors’ knowledge. Here, the author describes training the LoRA on analog photographs. I do not know what permissions accompanied those images. What I can see is the result: their photographic qualities help make an invented scene look convincing to me.

I might remove a small mark from the skin of someone I had photographed. Here, a similar detail makes me more willing to believe that someone was photographed at all.

As a photographer, I find some reassurance in why people still want photographs of their own lives. A couple asking for pictures together, or getting married, wants a record of themselves and of something they have lived. However convincing the generated café scene may be, it cannot meet that need.

I feel less reassured when I think of fashion and advertising, where a commission may depend on creating a convincing scene. The article raises the possibility that a brand could obtain photographic-looking images without hiring a photographer, renting a studio or organising a shoot. Looking at the generated portrait, I wonder how often a photographer will remain part of that process. Seeing a result I cannot visually distinguish from a photograph makes that concern more immediate.