benchrAll capabilities Image, audio and video
Generate an image with legible text inside it
The failure everyone remembers from early image models - garbled lettering - is documented as a supported capability now, for infographics, menus, diagrams and marketing assets.
DocsOfficially documentedTestNot tested by benchr
Gemini API
Overview
- Generates and edits images on gemini-3.1-flash-lite-image, gemini-3.1-flash-image, gemini-3-pro-image and gemini-2.5-flash-image.
- Keeps editing conversationally across turns through previous_interaction_id, so a correction does not restart the image.
- Accepts up to 14 reference images on Gemini 3.1 Flash Image, which is what makes style and product consistency possible.
How it works
- 01Write the exact copy in quotes in the prompt and say where it sits. Do not leave wording to the model if it must be correct.
- 02Attach reference images for brand assets rather than describing them.
- 03Correct in the same conversation instead of re-prompting from scratch, so the composition survives.
- 04Check the resolution ceiling before you pick the model - the Flash Lite image model supports 1K only.
Limits
- All generated images carry a SynthID watermark.
- gemini-3.1-flash-lite-image supports 1K resolution only.
- Reference images cap at 14 on Gemini 3.1 Flash Image.
- Legible text is described as a strength, not a guarantee - long strings and small type remain the weak case.
Verification
benchr test runs
benchr has not run this capability. No test has been published.
benchr test runs →
Evidence
Sources
- Google - Gemini image generation: model ids, legible text, 14 reference images, SynthID Tier 1 · September 1, 2026
Not stated by the source
- Character accuracy rates for text rendering
- Whether the watermark survives downstream editing
Source-check history
- September 1, 2026DocsOfficially documentedFirst logged. Model ids, the legible-text claim, the 14-image reference cap, the 1K limit on Flash Lite and the SynthID watermark read from the vendor page.