How to prompt GPT Image 2.5 for film work: Flare vs Sunburst, ten aspect ratios, quality levels, a reusable structure and three worked examples with tips.
GPT Image 2.5 is OpenAI's image model, and Cinemagiq runs it in two lanes: GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. For film work it handles concept art, storyboard sheets, character and expression sheets, posters, title cards and first frames for image-to-video. Its ten aspect ratios mean a frame can come out in the exact shape a deliverable needs, from a square expression sheet to a 21:9 scope frame.
GPT Image 2.5 takes a written prompt and a handful of settings:
For a sheet with several panels, say the layout first, then describe each panel in order. For a single frame, open with the medium or shot, then the subject and setting.
[Image type or layout]: [subject, or each panel numbered in order], [setting], [light], [style or medium], [labels or captions, with the exact words if they matter].
For a single frame, skip the numbering. For a sheet, make the count and the grid agree: six panels means a 3 by 2 or 2 by 3 grid.
Made in Cinemagiq. The exact prompt is under each one.
PromptA storyboard sheet with three numbered frames in a row for a coffee commercial: 1) close-up of coffee beans falling, 2) espresso pouring into a white cup, 3) a hand lifting the cup toward the camera. A short caption under each frame.
PromptWatercolor illustration of a quiet canal town at dawn, small wooden boats, laundry lines between the houses, soft pastel light, loose brushwork, no people.
PromptCharacter expression sheet: the same freckled girl in a yellow raincoat with six expressions (happy, sad, surprised, angry, scared, laughing) in a 3 by 2 grid, consistent face and outfit, animated film style, each expression labeled.
Example 1: a storyboard sheet on Sunburst. The layout is stated before any content: "three numbered frames in a row". Each frame then gets its own numbered beat with a shot size or action, "1) close-up of coffee beans falling, 2) espresso pouring into a white cup, 3) a hand lifting the cup toward the camera", so the sheet reads as a sequence. "A short caption under each frame" asks for text without dictating it. This one runs on GPT Image 2.5 Sunburst, the precision lane, because a multi-panel sheet has many parts that each need to land.
Example 2: a watercolor on Flare. The medium leads, "Watercolor illustration", and the prompt then describes how it is painted, "loose brushwork", not just what is painted. Concrete details ("small wooden boats, laundry lines between the houses") give the scene a sense of place, and "no people" keeps the frame empty. A look study like this suits GPT Image 2.5 Flare, the fast lane, where you run many variations.
Example 3: an expression sheet on Flare. The character is fixed in one phrase, "the same freckled girl in a yellow raincoat", and repeated as an instruction: "consistent face and outfit". The six expressions are listed in parentheses and the grid matches the count, "in a 3 by 2 grid". "Each expression labeled" turns the image into a usable reference sheet, and the 1:1 ratio fits a square grid.
All three examples ran at medium quality.
{{placeholders}}, for example an expression sheet with {{character}} and {{style}}, so every character gets the same layout.They are two lanes of OpenAI's GPT Image 2.5 in Cinemagiq. GPT Image 2.5 Flare is the fast lane, a good fit for exploring looks. GPT Image 2.5 Sunburst is the precision lane, for exact edits and frames where the details have to land. Both offer the same ten aspect ratios and quality levels.
Ten: 1:1, 4:3, 3:4, 5:4, 4:5, 16:9, 9:16, 3:2, 2:3 and 21:9. That covers square sheets, portrait posters, widescreen frames, vertical social cuts and a 21:9 scope frame, so the image comes out in the shape the deliverable needs.
GPT Image 2.5 offers five quality levels, from low up to max. Explore compositions at a lower level, then run the prompt you have chosen at a higher level for the frame you will keep, print or animate. All three examples in this guide ran at medium.
No. Camera Angle Control works only on the Google image models, the Nano Banana family. On GPT Image 2.5, write the angle into the prompt itself, for example 'low-angle shot' or 'top-down view', near the start where it frames the rest of the description.
Run GPT Image 2.5 and the other leading models in one AI-native workspace.