This document covers everything you need to replicate the workflow from the video, including the Gemini Gem, source files, and the storyboard technique discovered during testing.
The JSON Image Creator Version 3 is an updated Gemini Gem built for structured AI image prompting. Instead of typing a description directly into an image model, you use the Gem to generate a JSON-formatted prompt first. That JSON output gets pasted into GPT Image 2 to create the image.
The reason JSON works better here is structure. A plain text description leaves room for the model to guess. JSON defines the fields before generation starts, which tends to produce more consistent, controlled results.
A paid ChatGPT account is required for the storyboard feature. As of the recording date, GPT Image 2 is available in the US but availability may vary by region. The Gemini Gem can be used with any Gemini account.
The JSON Image Creator Version 3 Gem is included in the Notion doc download. You can either use the ready-made Gem directly or use the source files to create and modify your own version.
When you paste a plain-language image prompt into the Gem, it returns a JSON-structured version of that prompt, complete with camera type, aspect settings, and meta tokens baked in. Meta tokens are specific descriptors that push the image toward more cinematic and detailed results without you having to write them manually.
One thing to know: if you use the same Gem chat session across multiple prompts, the Gem will occasionally generate an image instead of returning the JSON. If that happens, go back and edit the message to include the phrase "a prompt for" before your description rather than reprompting fresh. Reprompting wastes tokens. Alternatively, start a new chat in the Gem.
GPT Image 2 has a built-in aspect ratio button on every generated image. You can switch from the default square to 16:9 landscape for YouTube-style images, 9:16 for TikTok and short form, or a range of other formats. The model re-renders the image in the selected ratio while preserving the content. This is fast and works well.
