Old words for new images

Gesso is the primer a painter puts on a canvas before the first brushstroke: chalk and glue, built up in thin coats. You never see it in the finished painting, but it decides how the paint sits and how long the work lasts. I named the site after it because the part of a generated image I care about most happens before the image exists, in the request.

Gesso Club is a curated archive of AI images. I choose them the way I’d judge a painting or a photograph, by composition, light, color and proportion. When I can find the prompt, model and workflow behind an image, they go in its record. When I can’t, the record says so. The site also has a Builder, a tool that writes prompts in the language photographers use. This post explains why the archive and the Builder belong together.

What the camera used to decide

Every photograph is the sum of dozens of decisions: lens, light direction, exposure, film stock or sensor, how the print was made. A camera settled many of them before the photographer pressed the shutter. One body, one lens, one roll of film with a fixed range of tones. Photographers developed their taste inside those limits, and often chose the limits on purpose.

An image model comes with none of them. Nothing about light, lens or grain is fixed in advance, so each of those decisions is made either in the prompt or by the model. A prompt like “a woman on a beach at sunset” leaves nearly all of them open, and the model fills the gaps with its defaults, in roughly the same way for everyone who asks. That’s where the recognizable AI look comes from: waxy skin, soft even light on the face, a glowing rim around the subject, teal in the shadows. Most of that isn’t the model’s ceiling. It’s what the model does when nobody tells it otherwise.

So cheap images aren’t the problem. If people had first met photography through Snapchat filter selfies, plenty of them would have concluded that photography can only make slop. The problem is unmade decisions. A generated image is only as good as the decisions the person asking actually made, and how precisely they could state them.

Words that name causes

Photographers ran into the precision problem long before AI did. Printers, cinematographers and colorists needed to tell each other exactly what to change, and over nearly two centuries they built a vocabulary for it. What makes that vocabulary useful for prompting is that most of its terms name a cause rather than an impression.

“Cinematic” names an impression. It could mean a widescreen frame, anamorphic flares, a teal-and-orange grade, shallow focus or low-key lighting, so a model asked for “cinematic” tends to apply some blend of all of them. It asks for every movie at once. “Rembrandt lighting” names a cause: the key light placed high and to one side, so the shadowed cheek catches a small triangle of light. It changes one thing in the picture, and you can check whether you got it. Halation works the same way. It’s the red glow around bright highlights that appears when light passes through the film, bounces off the base and exposes the emulsion a second time. Name the process and you’re asking for a specific effect you can verify.

This is also why the argument isn’t film versus digital. A clipped digital highlight and film’s gradual rolloff into white are both choices you can name, and sensor noise dates a picture as reliably as Kodachrome’s colors do. What matters is being able to name what you see.

Where the old words stop working

It would be convenient if every model understood this vocabulary. Many don’t. I’ve asked models for “shot on an ARRI ALEXA 35”, “raking light” and “dutch angle” and gotten back images that looked like the words weren’t there.

One likely reason is how models learn what words mean: from the captions paired with their training images. OpenAI trained DALL-E 3 mostly on captions written by another AI model (95% synthetic), and many labs have since done the same. A captioning model that describes a picture as “a woman on a beach at golden hour” will never teach the image model what halation is, however many of its training images show it. A term only steers a model if the captions used it.

So these words don’t come with a guarantee. They’re the most precise language we have, and whether a given model understands a given term is something you have to test. That’s why I built the Filmmaker’s Benchmark, which measures how prompts for film stock, lighting and scan treatment actually change an image’s color, contrast and texture. The hardest lesson from it was about noise. The same prompt run twice can produce noticeably different images, so a word only counts as working if its effect is bigger than that run-to-run variation.

Why the archive keeps the record

This is where curation comes in. A strong image with its prompt attached is evidence: it shows which words a particular model responded to, and how. A strong image without its prompt is still worth looking at, but you can’t learn how to ask from it. That’s why every record in Gesso Club says what’s known about how its image was made, including when the answer is “nothing”. Many images reach me without their prompts, because many people don’t publish them. My own generations, posted as User Zero, always carry the full prompt, model and workflow.

What the Builder does

The Builder applies the same idea to writing prompts. Every control writes a phrase a working photographer would recognize. Set the color temperature to 3200K and the prompt reads “3200K tungsten warmth”. Pull the shadow end of the tone curve down and it writes “crushed blacks”.

Two rules follow from the argument above. A dial left at its neutral middle writes nothing, because it’s a decision you’re deliberately leaving to the model, and filler words only compete with the ones you chose. And the same settings always produce the same prompt, word for word. That makes the Builder useful for testing: if two images made from the same board look different, the difference came from the model or the seed, not from a change in wording.

None of this is nostalgia for film. The camera used to make many of an image’s decisions for you. A model makes them too, but by default, and the same way for everyone. The alternative is to make them yourself, in words precise enough for the model to follow, and to keep a record of which words it actually followed. Photography already has the words. The glossary below is where I’d start.

Appendix: the working vocabulary

The words the Builder writes, and what each one asks of an image. They’re the terms most likely to steer a model, but check them against the model you’re using, because a model only knows the words its training captions taught it. Light comes first, because every other decision depends on it.