Tested H3-Image on a text-heavy sign: 4 of 4 generations got every line right
retained text
Wanted to see how MiniMax H3-Image (the new open-weights one) handles a sign with real copy on it, not just a two-word logo. One prompt, 4 generations, 3:4 at 2K, no edits or retries. All 4 got every line right, including "35mm B&W", both en dashes in "MON–FRI 12:00–18:00" and the tiny footer. What changed between runs was the design (serif vs sans, pinned vs framed). The text stayed exact. Two of the four attached. Prompt: \> Design an original wall-mounted information board for a fictional film-processing shop, shown straight on against a muted terracotta wall. Use cream paper, charcoal lettering, narrow dark-red rules, generous spacing and no photographs. The sign itself occupies most of the frame. At top show these exact lines: "FRAME ROOM" and "FICTIONAL STUDIO — SAMPLE INFORMATION". Below, arrange three paired service-and-inclusion rows under the headings "SERVICE" and "INCLUDED". Separate the rows with narrow dark-red horizontal rules and generous spacing, without a rigid boxed spreadsheet grid. Row 1: "35mm COLOR" | "DEVELOP + SCAN". Row 2: "35mm B&W" | "DEVELOP + SCAN". Row 3: "120 FILM" | "SCAN ONLY"; keep both phrases on the same horizontal row. In a separate lower panel print exactly: "COLLECTION" then "MON–FRI 12:00–18:00" then "Bring your claim slip." Put "No prices shown — sample layout only." in a small readable footer. Preserve each phrase in its assigned row; do not add prices, claims, addresses or extra services. Crisp readable type, realistic flat paper texture and one soft shadow beneath the frame. Portrait image, 3:4. Has anyone pushed it further, like a full menu or a dense infographic?
document fetched 2026-11-11 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xsrnb0/tested_h3image_on_a_textheavy_sign_4_of_4