AI Image to 3D Model: Connecting a Multimodal Workflow with a Blender Browser Viewer
From a single prompt to a browser-viewable 3D Fabergé egg in 18 minutes, using GPT-4o and Blender.
Developer Simon Willison demonstrated a complete multimodal AI workflow: generating a Fabergé egg image with ChatGPT Images 2.5, feeding it into a coding agent (Codex with GPT-4o Astra) that autonomously wrote and executed Blender scripts using a predefined skill file, producing a complex 3D model with 783K vertices and 17 materials in under 18 minutes. A vibe-coded browser-based .blend viewer made the result accessible to anyone without installing Blender. The core insight: AI agent power comes from chaining image generation, coding, 3D rendering, and browser tools — not from any single model's capability.
Developer Simon Willison recently shared a fascinating experiment: starting from an AI-generated image, he used GPT-4o Astra combined with Blender to produce a 3D model that can be viewed directly in a browser. This project links image generation, coding agents, and 3D modeling into a single pipeline — a compelling demonstration of what today's multimodal AI toolchains can accomplish.
Where It All Started: A Fabergé Egg as the Creative Seed
The experiment began with a fun idea. As a fan of the Imperial Russian Fabergé eggs, Willison wanted to create new eggs themed around pop culture. He turned to ChatGPT Images 2.5 and entered this prompt:
Generate a photo of a faberge egg that's themed after the TV show Pluribus - research first
You might have noticed the "research first" instruction — it asks the model to understand the show before generating anything. The result was impressive: a half-open yellow egg revealing a miniature scene inside, featuring the character Carol dressed in black and facing a crowd of eerie white-faced conjoined figures, set against cacti and the desert atmosphere of Albuquerque, New Mexico, with a Pluribus nameplate on the stand. Willison called the first attempt "honestly pretty great."
From Image to 3D Model: The Coding Agent Takes Over
The real power of the toolchain emerged in the next step. Willison pasted the generated image into Codex running GPT-4o Astra (high), then gave it this instruction:
Use your blender local skill to create a blender model of this faverge egg
The key here is the predefined "skill" file — a capability wrapper that teaches the coding agent how to interact with Blender. Willison had developed this local Blender skill through prior experimentation. With it in place, the agent could autonomously write and execute Blender scripts to complete the modeling work.
The entire process took 17 minutes and 51 seconds, producing several .blend files. From image to 3D model, the whole pipeline was driven by the AI agent — humans only needed to provide a high-level statement of intent.
The Blender Browser Viewer: Preview 3D Models Without Installing Anything
Having a .blend file is one thing — actually opening it is another challenge for most users. So Willison brought in a small experimental tool he had previously built through "vibe coding" and integrated it into his toolkit: the .blend URL Viewer.
The elegance of this browser-based viewer is simple: give it a URL pointing to a .blend file, and you can view the rendered 3D model directly in your browser — no Blender installation required. It can even parse file links from GitHub via the jsDelivr CDN proxy.
The viewer's info panel reveals the impressive scale of the generated model:
- File size: 7.2 MB
- 387 meshes, 1,099 curves
- 783,764 vertices, 1,446,560 triangles
- 17 materials
The Final 3D Model: How Faithful Is the Detail?
The result is a gold, jewel-encrusted Fabergé-style egg: the top opens to reveal a miniature desert scene inside — green cacti, brown rocks, and a small figure in a black suit at the center. The shell is decorated with green gems, pearls, intricate scrollwork, a ring of smiling white egg-shaped faces, and oval medallions painted with cacti and desert landscapes. The egg rests on four gold legs atop a black marble base engraved with the PLURIBUS nameplate.
The viewer honestly notes its limitations: "Materials and text are approximate; unapplied modifiers are omitted" — a candid acknowledgment of technical constraints.
What This Multimodal AI Workflow Reveals
Beyond the novelty of the egg itself, the real value of this experiment lies in demonstrating a complete multimodal AI workflow.
Reusable Encapsulation of Capabilities
By defining a skill file, Willison codified the complex knowledge of "how to operate Blender" into a reusable unit that coding agents can invoke repeatedly. This is central to making AI agents practically useful — crystallizing domain knowledge into composable capability modules.
Natural Handoffs Across Modalities
The image generation model handles "creative visualization," the coding agent translates "image into 3D scripts," and the browser tool handles "presenting the result." Each component plays its role, and humans only need to provide intent at the key junctures.
Building and Reusing Your Own Tool Ecosystem
Willison has long accumulated small utilities through vibe coding, and the Blender browser viewer is a perfect example of a ready-to-use tool assembled on the fly. When someone has built up a rich enough personal toolset, new experiments can be rapidly snapped together.
For developers, this experiment sends a clear signal: the power of AI agents lies not in how smart any single model is, but in how well you can chain together generation, coding, and rendering capabilities into an end-to-end automated workflow. The distance between a single natural language prompt and an interactive 3D artifact is shrinking fast.
Related articles
Insufficient Source Material to Generate a Valid Article
The provided source material is a single unrelated tweet with no AI or tech relevance — insufficient to support a complete, valid technical article.
Insufficient Source Material to Generate a Valid AI/Tech Article
This source material is a tweet about the ages of Underworld members — unrelated to AI or tech, and insufficient to support a full article.
Insufficient Material: Unable to Generate a Valid AI/Tech Article
The provided material is a condolence tweet about a San Diego mosque attack — unrelated to AI/tech and too limited to generate a valid technical article.