Built from research, not guesswork. The research behind Brah's live 3D and whiteboard scenes.

Why AI drawings often look like toys

Ask an AI to build an object in 3D and you often get a stack of boxes and cylinders. The idea is right and the shape is wrong. A toy picture of a heart explains less than one that looks like a heart.

Researchers saw the same thing. In LL3M, AI agents write Blender code to build 3D objects. Without reference material, the coding agent often fell back on a few straightforward primitives and skipped the more complex operations.

The AI describes the scene. Brah draws it.

The AI does not draw pictures, and it never runs code on your computer. It writes a description of the scene: the steps, and in each step the parts, the labels and what moves. Brah checks that description, then its own engine draws it next to your cursor.

The scene is data: a list of steps, each with operations such as add a part, label it or animate it. Brah checks the data and renders it with three.js. In all three papers below, the AI writes Python that runs in Blender. Brah keeps the readable, step-by-step build, as data it can check before anything appears.

Real shapes from general parts

Brah has no ready-made models of hearts or bikes. The AI builds each object from general parts:

• swept bodies that change size along their length
• flat outlines pushed out to a depth
• turned forms, like a vase on a lathe
• tubes for pipes, rods and wires
• custom meshes for anything else
• repeats, for rows and rings of the same piece
• cut-away views, to show what is inside

Each part takes the colour of the real thing's finish, such as a brass lamp shade. The whiteboard keeps Brah's marker colours.

The parts an AI has to work with decide whether the result looks real or like a toy. When LL3M gave its AI Blender's documentation, it used about five times as many complex operations across 17 test objects, and the models had more detail: rods and fins on a windmill, curved threads on a lantern.

Plan the parts, then build

Before it draws, the AI lists the object's main parts, their sizes and where each one joins the next. Then it picks each part's kind by its real form. A round, turned form is a lathe. A bent pipe is a tube. A part is a box only when the real thing is a box.

The papers split the work the same way. LL3M has a planner agent that breaks a task into parts (its example: a chair becomes legs, a backrest and a seat) before a coding agent builds each one. 3D-GPT has separate agents to pick the functions, fill in missing detail and write the code. SceneCraft first plans a scene graph of how the objects sit in relation to each other, then turns those relations into layout constraints.

The right guide for each scene

Brah keeps a short guide for each kind of scene: 3D objects, whiteboard drawings, steps, motion and photos. The AI loads only the guide for the scene it is drawing.

This follows LL3M, where the reference material changed how much of the shape language the AI used. LL3M searched a store made from 1,729 Blender documentation files to find what each part needed. Brah's guides are much smaller. Loading one at a time also keeps the AI's context small. An AI that cannot load guides gets them all at once.

A method, not a template

The guides teach how to build. They never include a finished example object, because an AI that copies an example makes every drawing look the same.

3D-GPT saw this happen. When its modelling agent had to guess values for a shape, it tended to use defaults or copy values from the examples in its prompt, and the results were less varied.

The AI checks its own drawing

After a 3D scene shows, Brah draws it again, finished, out of sight, from the front, a side and the top. It sends that one picture back to the AI that drew the scene. The AI checks it once and fixes only what is clearly wrong, such as a part that floats. You do not wait for this: the scene is already on screen, and the check runs while Brah talks.

The check works with Claude, ChatGPT and the Anthropic API. Local models and Hermes do not get it. Whiteboards are not checked yet, because their strokes draw in over time, so a picture taken at once would be blank.

LL3M and SceneCraft also show the AI what it made, in different ways. In LL3M, Claude writes the code, and a separate vision model, Gemini, looks at five renders for problems such as chair legs that do not touch the seat. A verification agent then renders again to check the fixes. SceneCraft is built on GPT-4V: a reviewer compares each render with the description and revises the script, round after round. Brah asks the same AI that drew the scene, from three views, once, so an answer cannot loop. The picture takes about 0.04 to 0.8 seconds to draw.