NemoClaw Knowledge Wiki

Nano Banana 2: JSON Control for Precise AI Image Editing in Gemini

4 min read

Nano Banana 2: JSON Control for Precise AI Image Editing in Gemini

Clip title: Nano Banana 2: The JSON Control Hack Author / channel: renderdrop URL: https://www.youtube.com/watch?v=uQc4TGhvDHc

Summary

The video introduces a “game-changer” for AI image editing, addressing the common frustration of AI models “hallucinating” or ruining an entire image when a user attempts a minor modification. The solution presented involves using the JSON (JavaScript Object Notation) code format within Google’s Gemini app, powered by the Nano Banana 2 model (specifically Gemini 3.1 Pro). This method grants users “ultimate control” by breaking down an image into a structured code format, allowing for granular manipulation of individual elements, camera angles, lighting, and other properties, thereby preventing unintended alterations and maintaining the original image’s integrity.

The tutorial demonstrates this powerful technique through five distinct use cases. First, for Colors & Materials, the video shows how to analyze an interior design image to generate a detailed JSON code. Users can then precisely edit specific objects’ colors (e.g., changing an armchair from cream to light blue) and materials (e.g., velvet or oak wood) by modifying the JSON, with the AI executing only the specified changes while preserving all other elements and the scene’s perspective. The second use case tackles more complex Object/Furniture Swaps, illustrating how to replace an entire armchair with a new one. This involves generating JSON for both the existing scene and the new object, then cleanly merging them (using a separate Gemini chat to avoid errors) to ensure the new furniture is perfectly integrated, including correct positioning and realistic shadows, even if its original orientation was different.

The third and fourth use cases delve into environmental and technical aspects. Weather & Lighting demonstrates transforming a sunny interior into a moody, rainy day, or a golden hour scene. The key here is using a specialized JSON prompt focused on lighting, weather, and shadows, and understanding that precise wording is crucial to prevent the AI from making unwanted changes (like removing curtains to “show” rain). For Camera Perspective, a notoriously difficult task for AI, the video showcases extracting a complex fisheye perspective from one image’s JSON and applying it to a different interior scene. Despite the AI needing to “hallucinate” the expanded edges, the core perspective transfer is achieved flawlessly, demonstrating impressive control over photographic properties.

Finally, the video explores Changing Text & Logos, a challenge many AI models struggle with regarding consistency and texture. By generating a JSON focused on typography and branding elements, the presenter successfully changes “THE BREAD OF LIFE” (composed of toast letters) to “SUB TO RENDER DROP,” accurately replicating the bread texture and text structure. A more advanced example replaces a complex Louis Vuitton logo, formed by boats in the sea, with a custom “R” logo, again showcasing the ability to maintain the intricate object arrangement and overall coherence. The conclusion emphasizes that this JSON-based method provides an unprecedented level of detailed control, mitigating AI’s tendency to generalize or hallucinate, and encourages viewers to try the provided prompts to experience this precision themselves.

Graph View

Backlinks