Testing all five official GPT Image 2.6 prompting tips: Sunburst vs Flare
Fresh-Resolution182 · reddit · 2026-11-10
The author empirically tested five tips from the official GPT Image 2.6 prompting guide on the Sunburst and Flare variants — 12 real outputs at matched size/quality/complexity, with timing and cost logged (total $0.212692 billed, 7m44s combined latency).
Findings:
- Exact poster text: quoting copy verbatim ("NIGHT BLOOM", "FRIDAY 8 PM") and banning other words — both spelled everything correctly; Sunburst cleaner, Flare more illustrative.
- Camera & lighting specs: 85mm lens, bottle-height camera, warm key from left, cyan fill from right, leaf shadows — both complied; Flare pushes contrast, Sunburst stays neutral.
- Invariant lists for edits: explicitly listing shape, cap, color, position, scale, angle, reflections, edges as unchanged far outperformed vague "replace the background".
- GIF loops: specifying 8 frames, locked camera, 140ms/frame, matching first/last poses produced real animated GIFs from both.
- Short vs structured prompts: one-line prompts let models invent freely; structured specs (object/palette/framing/light/exclusions) converged both on the intended still life.
Takeaway: start short for exploration, switch to labeled structure when composition matters. Flare was faster on three of four simple stills but slower on edits and GIFs — choose by task, not brand.
More from Multimodal
- Researcher Uses Claude to Craft a Convincing Fake Pharma Ad Over One Weekend — andrey_kurenkov · 2026-11-10
- Delay Lama mod now does beatbox-style vocal percussion — sebkrier · 2026-11-10
- Redditor's AI-Animated Book Was a Success — Ok-Revolution-2453 · 2026-11-10
- Claude persona "The Tidewright" spins out stunning aerial-view landscapes — repligate · 2026-11-10
- Fable 5.2 Squashes a 3D Renderer Into a Working Game Engine — sebkrier · 2026-11-10
- Seedream 5.5 vs GPT Image 2.5/2.6: photorealistic Chinese-text rendering face-off — videorouter · 2026-11-10