Production-Resolution AI Retouching

The Challenge

"Replacing the full foreground of a production file without a seam."

Jordan & Dani's Subaru frame had great light and a great moment. It also had a gravel road across the foreground that pulled the eye, and a blurred finger creeping into the lower right corner. The fix was to replace both with native grass and plants.

• Production Resolution: The file is over 8,000 pixels wide, and current AI models top out at 4K per render. The area being replaced runs the full width of the frame, so it had to be built in at least two pieces.
• Consistency Between Renders: Each piece had to match the others in grass type, light, depth of field, and color. AI can produce a great result once and then struggle to repeat it.
• Why this was not a simple retouch: Most AI tests you see online are single prompts on low-resolution images viewed on a phone. This had to hold up at 100% on a campaign file.

The Objective & Success Criteria

"A foreground that looks photographed, not generated."

• Visual Fidelity: New grass and plants that read as real at full resolution, with natural focus falloff that matches the lens.
• Seamless Integration: No visible seams between renders and no change to the vehicle, the talent, or the light.
• Tool Evaluation: Find out whether OpenAI's newest image model, GPT Image 2.5 Sunburst, can replace Nano Banana 2 in our production workflow.

Our Approach

"Two models, one prompt, judged at 100%."

Both models ran inside Photoshop through Astria's plugin, so each got the same file, the same prompt, and the same workflow:

"Remove the gravel road and the blurred finger in the lower right corner, and replace them with native grass and plants."

• GPT Image 2.5 Sunburst: At its max quality setting, Sunburst got both halves on the first try, with color close to the original. The renders were usable starting points that could be retouched into place.
• Gemini 3.1 Flash Image (Nano Banana 2): Using the same prompt, plus an instruction not to change the focus or the color, Gemini took about 10 renders to produce two usable ones. It also shifted the grass from green to a dry, golden tone.
• The 100% Review: From a distance, Sunburst looked like the clear winner. Zoomed in, its image fell apart. The plants didn't read as plants; the detail had a crunchy, overprocessed AI texture, and the depth of field was flat. Gemini's blades read as real grass, and its focus falloff looked like it came from a real lens.
• Retouching Into Place: We chose the Gemini renders, corrected their color to match the original, and composited the pieces into a single seamless foreground.