An experiment: recreate JSON-prompted closed model image in Ideogram 4
A user on Reddit conducted an experiment to recreate an image generated by a closed model using Ideogram 4. The process involved summarizing the original JSON prompt into natural language with an LLM, then using that prose and the original image to generate a new JSON prompt for Ideogram 4 with the Gemma 4 12B QAT model. The user then adjusted bounding boxes and generated the final image, sharing the workflow for others to replicate. AI
IMPACT Demonstrates a novel cross-model image generation technique using LLMs for prompt translation.