On 13 September I asked for every headshot the ARCHV desks use to be remade the way the Shelton x Rodman art was made: same model, same prompt. Both were already the same.
The Shelton x Rodman carousel went out on 9 September, built on illustrated portraits of Ben Shelton and Trinity Rodman. With 44 shares it is the most-shared post on its account, and on 13 September it became the house rule for carousels on two ARCHV accounts, with an illustrated portrait across the full width of every slide, at least 1080 pixels deep in a 2700-pixel frame. Two days earlier, on 11 September, I had fixed the portrait prompt. It is 22 words, and the order said nothing could be added to it, including by the platform's own prompt enhancer. The desks generated 46 portraits against that order on the first day, and the written order told them to read the prompt back from each job to confirm nothing had been appended.
By the 13th I did not like some of the faces, and the bar I named was the Shelton and Rodman art.
What was different between the two sets of jobs?
The image platform keeps the parameters of every job it runs, so the record answers the question. Put a Max Verstappen portrait job from 11 September beside an Alexander Zverev job from the rebuild, made to the Shelton x Rodman settings.
Both ran on gpt_image_2 with the same 22 words as the prompt. The Verstappen job asked for a 1:1 image at 2k resolution and medium quality. The Zverev job asked for 3:2, 4k and high. The output follows from those three fields: 2048 by 2048 pixels, about 4.2 megapixels, against 3504 by 2336, about 8.2. So does the price. My credit ledger charges 2 credits for the first and 11 for the second.
The order written down on 11 September carried the Shelton x Rodman prompt word for word and, beside it, settings of 1:1, 2k and medium. Only the prompt matched the original job.
I cannot tell you which of the three made the difference. They changed together, and the rebuild moved all three back at once.
Why a square portrait fought the layout
The shape was a problem before the quality was. A carousel slide on these accounts is 2160 by 2700 pixels, and the portrait crosses it edge to edge as a band at least 1080 pixels deep, and 1540 on the opening slide. A 3:2 image is already close to that shape. A square one has to be cropped hard or made wider.
On the morning of 13 September my ledger shows a run of outpaint jobs, which paint new canvas around an existing image, and one upscale. Then, at 07:51 Eastern, the first jobs at the Shelton x Rodman settings went through. The rule written that day bans both repairs for faces. A portrait that falls short is generated again at the recipe settings or the post waits, and an old face is never widened or enlarged to fit.
The reference photo is part of the recipe
The model repaints what it is shown. The first rebuilt Wayne Rooney came back with the Champions League starball from the background of his 2009 reference photo painted in behind him. Painting it out of the finished portrait left a visible patch. The version that shipped came from a reference whose whole background had been blurred before generation, and the starball never appeared.
So crests, sponsors and maker's marks now come off the reference photo before the job runs, and never by adding words to the prompt. A reference whose face is too small or too soft to carry a likeness is replaced rather than enlarged.
What each bank row keeps now
Each headshot bank row made since 13 September names the job id and the id of the cleaned reference as it was uploaded. The settings are checked against the job record itself rather than against a note about it. The cleaned reference is stored because it is an input to the job, and a fresh download of the original photo is a different file.
The rebuild was eight jobs and 88 credits, for seven faces. The eighth was the Rooney with the starball.
Between the Verstappen job and the Zverev job, the prompt field differs by one trailing space.