GPT Image 2.5 guide: Flare hits 22s for high quality vs 177s, plus character consistency workflow

Fun_Walk_4965 · reddit · 2026-09-10

Part 3/3 of a hands-on GPT Image 2.5 prompting guide covers choosing between Flare and Sunburst, quality settings, and character consistency:

Benchmarks: Same prompts, 1024 wide — Flare high text-to-image ran 22s vs 177s for GPT Image 2 high; Sunburst high took 48s. Sunburst max portrait: 115s.
Selection order: If your GPT Image 2 workflow passes, migrate straight to Flare; if not, get acceptable results on Sunburst first. Never compare on auto — set quality manually, tune one variable at a time.
Sizes: 1024 square up to 3840 long edge, edges multiple of 16, ratio ≤3:1, 655,360–8,294,400 pixels total; above 2560x1440 is experimental.
Multi-turn refine: pass output back with one change per turn, repeat key constraints; use EXACT verbatim phrasing for billboard text.
Characters: build one reference image, feed it into every scene with defining details repeated.

The full prompt library is open-sourced on GitHub (awesome-gpt-image-2.5-prompts).

More from Multimodal