Descriptions:
Nate Herk demonstrates a structured image generation pipeline that pairs Google’s Gemini 3.1 Flash Image Preview with Gemini 3 Pro running inside Antigravity, Google’s VS Code-based development environment. The core technique uses Gemini Pro as an upstream prompt architect: given a vague image request, it applies a reusable skill to convert the input into a structured JSON prompt specifying resolution, style, lighting, camera angle, and negative prompts before passing the result to the image model. The approach produces noticeably more consistent outputs than direct natural-language prompting, with all images in the video sharing a unified visual style generated through the same JSON structure.
The video highlights concrete improvements in the latest Gemini image model over its predecessor: faster generation, stronger prompt adherence, near-zero text misspellings in generated infographics (a persistent issue previously), and lower API cost per image. Side-by-side comparisons using hyperrealistic portraiture, advertising mockups, and game UI redesigns illustrate the quality-per-dollar argument. Google Search grounding can also be enabled to allow the model to reference real-time information during generation.
The tutorial walks through Antigravity project setup in full — configuring the gemini.md master system prompt, organizing output folders, and enabling the nanoban skill. The skill template is provided free to viewers. Practical for designers, content creators, and developers who want repeatable, cost-efficient image generation pipelines built on Google’s multimodal API stack rather than point-and-click tools.
📺 Source: Nate Herk | AI Automation · Published February 27, 2026
🏷️ Format: Hands On Build
