I had a weblog submit that wanted a header picture, and I figured I’d knock it out in two minutes with ChatGPT. Forty-five minutes later, I used to be nonetheless sitting there, regenerating the identical picture time and again, watching it give me a special model of the identical downside every time.
That’s the second I ended treating it like a merchandising machine the place you place in a immediate and an ideal picture comes out. I began treating it like a device I really needed to be taught. So I spent the subsequent couple of weeks testing this device, producing, modifying, failing, adjusting, failing once more.
What I landed on was eight issues that genuinely modified how I work with photographs in ChatGPT. None of that is about secret hacks or magic phrases. It’s about understanding what the device is definitely doing on the again finish and dealing with it somewhat than in opposition to it.
Here’s all the pieces I discovered, within the order I’d advocate tackling them.
Describe the picture such as you’re speaking to somebody who cannot see it
This sounds apparent, however it took me longer than I’d prefer to admit to really do it correctly. My early prompts have been quick and obscure, issues like “a woman at her desk working.” The photographs that got here again have been wonderful, however generic. They might’ve been inventory pictures from any website.
What modified issues was describing the scene the way in which I’d describe it to a good friend over the telephone who has zero visible reference. I began naming hair colour, the setting, the lighting, what’s on the desk, what the particular person is sporting.
If I had a selected image in my head, I’d put it into phrases somewhat than assume ChatGPT would guess my intent. If I did not have a transparent image, I observed the outcomes have been a coin toss, which made sense as soon as I thought of it. The AI has nothing to go on past what you give it.
Example immediate I used:
“A woman in her early thirties with short curly black hair, sitting at a wooden desk in a sunlit home office. She’s wearing a cream sweater and typing on a laptop. There’s a small potted plant and a coffee mug on the desk. Natural daylight coming from a window on the left.”
Don’t cram each element into one immediate
After I received higher at describing issues, I overcorrected. I began writing prompts that ran 5 – 6 sentences lengthy, making an attempt to regulate each single aspect. That backfired extra usually than it helped. The longer and extra layered the immediate received, the extra seemingly ChatGPT was to lose observe of one thing, often a small element buried in the course of the paragraph.
I discovered a greater stability by together with solely the small print that really mattered to me and letting the AI make cheap selections for all the pieces else. If the background colour wasn’t essential, I left it out as a substitute of specifying it. That gave the mannequin room to fill gaps sensibly as a substitute of making an attempt to fulfill ten competing directions directly.
Example immediate I used:
“A golden retriever puppy sitting in a grassy backyard, looking up at the camera. Bright afternoon light, candid and playful feel.”
Tell it how the picture needs to be made, not simply what’s in it
I used to solely describe the topic of the picture. I by no means mentioned something concerning the format, the temper, or the technical really feel of the shot. Once I began including that layer, my outcomes received noticeably extra constant.
Now I all the time suppose by way of just a few questions earlier than I write the immediate. Should this be a large shot or a sq. one? Does it must appear to be {a photograph} or extra like an illustration? Should the lighting really feel heat or cool? Should the colours be muted or vivid? Spelling this out gave ChatGPT a a lot clearer transient to work from, somewhat than leaving the “how” fully to probability.
Example immediate I used:
“A cup of espresso on a marble countertop, shot from a slight overhead angle. Square format, mushy pure mild, muted heat tones, seems like a way of life journal photograph somewhat than a…
