google/gemini-3-pro-image-preview/edit

What makes Gemini 3 Pro stand out

Gemini 3 Pro is a high-fidelity image-to-image editor built for realistic, production-grade results that preserve scene structure and layout. Using real-time reasoning, the model interprets intent, applies targeted adjustments, and maintains natural lighting, perspective, and material response. It excels at precise multilingual text edits and signage replacement without destabilizing composition. For demanding pipelines, Gemini 3 Pro sustains clean 4K output and consistent aspect ratios, enabling fast iteration with reliable continuity across variants. Gemini 3 Pro focuses on structure-aware changes rather than full-frame regeneration, minimizing artifacts and drift. Key capabilities:

• Structure-preserving edits: retains pose, layout, depth, and geometry to prevent unintended warp or re-synthesis.
• Lighting and realism continuity: matches shadows, reflections, and exposure for believable results.
• Multilingual typography: accurate rendering and replacement across languages with scene-aware placement.
• Localized control: region- and object-targeted changes instead of full-frame changes.
• High-resolution control: up to 4K with selectable aspect ratios for platform-specific delivery.
• Output flexibility: choose png, jpeg, or webp and create multiple variants per request.
• Stable iteration: Gemini 3 Pro keeps core composition intact across rounds of refinement.

Prompting guide for Gemini 3 Pro

Start with a base image and a concise prompt that names the subject, what to preserve, and what to change. Specify regions, lighting, style direction, and any text content (including language and font intent). Control outputs with resolution (1K, 2K, 4K), aspect_ratio, output_format, and num_images to explore safe variants before finalizing. For image-to-image tasks, Gemini 3 Pro responds best to explicit constraints like keep the subject, modify only the background, or replace the sign text with a specified string. Use short, direct instructions to prevent overreach by the editor. Examples:

• Preserve the model and pose; replace the background with a foggy forest, soft natural lighting.
• Remove the power lines; keep building edges and perspective intact.
• Replace storefront sign with "CAFÉ" in Spanish diacritics, white sans-serif, perspective matched.
• Add a wooden bench to the right of the subject; do not change the sky.
• Convert to editorial black-and-white grade; retain skin texture and clothing detail.
• Swap product label to Japanese text; keep bottle shape and reflections consistent. Pro tips:
• State preservation first: what must not change, then list targeted edits.
• Use spatial language: left, right, foreground, background, upper-right quadrant.
• Limit adjectives; prefer a few strong descriptors for style and lighting.
• Iterate in small steps; compare num_images variants, then refine.
• Provide clean, on-topic references; crop out irrelevant regions before prompting.