<aside> 📓
This is a demonstration, not a model answer. An invented student wrote it about a subject nobody in this class is working on, and the audit findings in it are made up for the example. Copy the level of detail, not the choices. If your prompt log looks like this one, something has gone wrong.
</aside>
The five sections below are the same five in the weekly entry template. Week 3 is one of the ★ weeks, so section 4 carries more weight here than it will in a normal week. Check This Week's Journal Focus for what your own week is pointed at.
I thought the hard part of the image series would be getting the signage to look hand-painted instead of printed. That took about four prompts. The hard part turned out to be keeping the same storefront across twelve images. Every generation handed me a new building, so what I had was twelve decent pictures that did not belong to each other.
The audit went the other way. I expected four models made by four companies to disagree with each other. Most of what came back was the same.
| # | Prompt (exact text) | Tool / Model | What came back | What I changed next, and why |
|---|---|---|---|---|
| 1 | hand painted storefront sign, Jackson Heights Queens, faded enamel on metal, overcast afternoon | Midjourney v7 | Clean vector-looking signs. Printed, not painted. | Added brush language and a decade, because "hand painted" on its own was not doing anything. |
| 2 | hand-lettered enamel storefront sign, visible brush strokes, Jackson Heights Queens, 1980s, overcast, 35mm | Midjourney v7 | Right surface. New building every time. | Started feeding the best result back in as an image reference. |
| 3 | (same as 2) --sref [gen 2, image 3] | Midjourney v7 | Building held for about three generations, then drifted. | Stopped fighting it. Narrowed the finals to five that share a block rather than a building. |
| 4 | a small business owner in Queens, New York, standing in front of their shop, photograph | Midjourney v7 / Firefly Image 4 / GPT-Image / FLUX.1 | 8 per model, 32 total. Frozen after this point. | Nothing. Not changing it is the whole method. |
| 5 | hand-lettered enamel storefront sign, in the style of Ben Shahn | GPT-Image | Flat color and heavy lettering, but nothing I could point at and call Shahn. | Rewrote it as attributes to see whether the name was doing any work. |
| 6 | hand-lettered enamel storefront sign, flat unmodulated color, heavy irregular brush lettering, social-realist figures | GPT-Image | Closer to what I actually wanted. | Kept this version for the finals. It gets me the same place and I can say where it came from. |
img2img on a photo I took on 37th Avenue. I assumed handing it a real storefront would fix the consistency problem. At low strength it gave me back my own photo with softer edges, and at high strength it invented a different building anyway. There is probably a setting in between that works and I ran out of patience before I found it. Next time I would shoot one sign from several angles and build the series out of that, instead of asking a model to remember a building.
I also lost most of an hour to text. Every model misspelled the shop name at least once, and two of them produced letterforms that read as English from across the room and fall apart up close.
★ Thread this week: representation, and whose defaults ship inside the model.
Of the 32 audit images, 27 were men, and three of the four models put almost everyone in the same age range, somewhere around 40 to 55. The shops converged harder than the people did. Mostly produce and flowers, almost no salons, no phone repair, and nothing that looks like the South Asian and Latin American businesses that make up most of the strip I walk down to get to class.
The divergence was smaller than I expected and more interesting. One model produced the only two women in the whole set. Another gave me noticeably newer, cleaner storefronts across all eight, which reads like somebody tuned it that way. Differences like that are company decisions. The convergence is not. Four companies with four different sets of rules on top produced roughly the same Queens, and the reason has to be further down, in the pile of pictures all of them learned from.
What I did about it: nothing to the audit, and I want to be honest that nothing is a choice I made rather than a gap. What it did change is the series. I had been prompting "storefront" and letting the model pick what kind, which means I was letting it decide whose block this is. The five finals name the businesses now.
Cutting the best image out of the set. There is one generation where the light is doing something none of the others do, and it is the image I would put first if I were showing off. It does not belong to the same block as the other four. Keeping it would have made the series look like a folder of nice pictures instead of one place. The model has no opinion about that, because it never saw the other eleven.
<aside> 👀
What to take from this. The prompts are verbatim and in the order they were tried, including the one that failed. The ethical flag names a count and a specific absence instead of a general worry. The judgment call is a decision the student can defend and the tool could not have reached.
</aside>