Prompting for Realistic Images: Portraits, Food, Nature and Interiors

In this March 2023 tutorial I experiment with prompts for realistic-looking images in four categories: portraits, food, nature, and architecture. I use Midjourney version 4 and discuss techniques that might also be tried in Leonardo.ai. Model behavior and interface controls have changed since this recording, and the examples are artistic experiments rather than controlled comparisons.

For portraits I start with a studio-style photograph of a woman, a light background, and a Kodak Gold 200 reference. The prompt gives the model cues about the setting and photographic look; Kodak Gold 200 is a film stock, not a camera. Naming a camera, lens, or film can suggest a style, but it does not mean a real camera was used or guarantee its optical characteristics.

Another approach is to reference a well-known photographer. I try prompts mentioning Annie Leibovitz and Richard Avedon, then inspect and upscale results. These names are style cues in the experiment, not evidence that the photographers created or endorsed the generated images.

I try descriptive terms such as photorealistic, hyper-detailed, sharp focus, and cinematic lighting, along with a Hasselblad camera reference. My first attempt produces pictures of a camera because I forgot to name the subject. Adding a woman in a blue dress makes the intended subject clear. Terms such as “8K” are cues; they do not guarantee a particular output resolution.

For a more natural portrait I add “natural skin texture” and look for wrinkles and texture instead of an oil-painting effect. I select a result to upscale. I also try a Polaroid look, inspired by another creator, and find that the blur and instant-film aesthetic make some examples seem more photographic.

I notice that sunglasses can hide unnatural-looking eyes in some generated portraits. I experiment with “straight teeth” and possible negative terms for unwanted teeth, although I cannot tell whether the wording is actually helping. Even the upscaled examples have imperfections, including areas that resemble paint or clay.

For food, one way to learn is to inspect public examples and the prompts behind images you like. I look at a detailed cake prompt describing layers, a plate, lighting, and texture. The generated cakes look convincing to me, though some areas look grainy. Borrowing ideas from a prompt and changing individual details can help explore their effects.

I often like simple prompts, but for food I find detailed descriptions useful: specify the subject, ingredients, surface, and lighting. In an upscaled cake image I look at crumbs, the fork, and the blurred background. I use “bokeh” as a cue for the appearance of out-of-focus areas; the term refers to the quality of that blur.

For nature I begin with a close-up ant and realism cues. The results contain fine texture, but I am unsure whether the anatomy is correct and notice an odd antenna. A detailed-looking generated animal is not necessarily an accurate biological reference.

I then ask for a parrot with high-quality and photographic cues, including a National Geographic reference. I like the apparent detail in feathers and beaks and upscale one result. I do not know whether writing 4K, 8K, or 12K changes the actual image quality; these words are not a substitute for checking the output dimensions.

A lion prompt combines wildlife-photography cues, a running subject, detailed fur, and “Unreal Engine 5.” I think the results are detailed, although there is grain and blur around some snouts. Naming a rendering engine is a stylistic cue in this prompt; it does not establish that the image was actually rendered in that engine.

For architecture I try an interior-design prompt describing a modern house and add a lighting cue. I want even illumination, but “global illumination” is a rendering concept involving direct and indirect light, not a guarantee of shadowless lighting. In a prompt, the model may interpret the phrase loosely.

A more specific interior prompt describes a sunken living room, wooden floors, small garden-facing windows, Bauhaus furniture, a high ceiling, and a pastel palette. I prefer the resulting examples, especially the light on the couch and the wood floor. The general lesson is to describe the scene and iterate, while checking whether individual details actually appear.

Across these examples, portraits are my personal favorite for apparent realism. Food, nature, and interiors can also produce appealing results, but the images are generated and may contain errors even when they look photographic.