My pick — the second column is what desktop adds, so the written-only state uses it. The ask becomes a 258px panel beside past work rather than a row beneath it, which is the desktop 1d arrangement the mobile version had to abandon; a homeowner sees the request and the evidence in one glance. Once renders arrive the ask panel gives way to the work: renders sit at 252×189 in a flex row, and past work persists rather than folding to a link — it fills whatever width the renders leave as a scrolling strip of 196×148 tiles running to the card edge, one card per column, with a glass arrow marking that more is reachable. The tiles are the same size in every state, so past work reads as one consistent band; the renders stay larger, which keeps them unmistakably the subject. Weakness: the states share no layout below the header, so a list mixing them looks less uniform than 1b.

One frame, two contents. The columns never move: renders left, past work right at 86px with a +13 overlay, so five submissions in a row line up exactly whether or not each has renders — the strongest option for scanning a list. The written-only state fills the left column with a dashed panel at the same 186px height the render grid occupies, so nothing jumps. Weakness: the dashed panel sits beside a full grid of past work, which reads as the PRO having done nothing for you — and the right column costs a quarter of the card in both states.

One strip, one header row, and the action rides in the header. Both states use the identical structure — a label row with something on the right, then one full-width strip — so only the label, the strip's contents and the image size change. Renders run at 392×294, the largest of the three concepts, which suits a surface whose whole job is judging a design. Weakness: the written-only state now shows past work at near-render size, which overstates its role; the label has to carry the distinction alone.