Ideogram 4 replicates closed model's photorealistic banana via reverse-engineered JSON prompts

A Reddit user known as GTManiK conducted an intriguing experiment: successfully recreating a photorealistic banana image originally generated by a closed, proprietary AI model using only open-source components. The original image was generated from a structured JSON prompt (part of the 'Nano Banana' series) hosted on a GitHub repository. To replicate it, GTManiK first summarized the original JSON prompt into natural language prose using an LLM, then fed that prose—along with the original image—into Google's Gemma-4-12b-QAT model. Gemma output a new JSON prompt specifically formatted for Ideogram 4, an open-source image generation model. The JSON was then passed through the KJ Prompt Builder node to fine-tune bounding boxes before the final generation.

The results were visually nearly identical to the closed model's output, highlighting the surprising flexibility of open-source tools when combined with clever prompt engineering. This experiment demonstrates that proprietary image generation systems can be effectively reverse-engineered without access to the original model weights. The entire workflow is documented in the linked GitHub repository, which provides the source JSON prompts and instructions. For AI artists and researchers, this opens up new possibilities for reproducing and building upon closed-model results, potentially reducing dependence on expensive or restricted APIs.