Multimodal chaining from concept to Blender model

The image is no longer the goal, but the input for the next step

Simon Willison shared a 3D model of a Fabergé egg for Blender on the social network X. There would be nothing unusual about this if the process of its creation didn't show where the AI workflow is heading. Willison first generated a 2D reference image using ChatGPT Images. He then inserted it as an input prompt for a model he refers to as "GPT-6 Astra" (in combination with Codex) and told it to generate structured code, specifically a Python script for Blender.

Shifting work from visualization to engineering

This workflow demonstrates that image generation is ceasing to be just a tool for illustrations. It is becoming an intermediate step. A developer or designer clarifies a concept using a fast text-to-image model, visualizes it, and then uses the ability of a vision-language model to read pixels and translate them into a mathematical representation. This pipeline directly yields code that standard industrial software can process.

Multimodal friction remains a limit

The reality check comes when we realize the need to chain several different specialized architectures together. The user still has to manually pass the output from one model to another. Although the models are great at generating scripts, creating an accurate and detailed 3D model purely from generated Python code often hits the limits of representing complex organic shapes using algorithms.

When will models start sending pixels themselves?

The key signal of future development will be native chaining. If platforms allow the image output from one function (DALL-E) to be passed directly to a coding agent within a single API call without user copying, the friction will be minimized. Tools like the Blender Python API will then act as a compiler for visual prompts.

Lilith's verdict

Human graphic designers won't be replaced by Midjourney prompt engineers, but by an automaton that turns concept art straight into compilable code for the factory.