Skip to content

MIDI

Generate multiple 3D instances from a single image

Visit official site

What MIDI is

The technique is described as an image to 3D scene approach using a multi-instance diffusion model. It produces several 3D instances with correct spatial relations from one image.

What you can do with it

  • Build a 3D scene from a concept picture
  • Compose several objects into one scene
  • Refine the generated scene texture

Key features

  • Multi-instance attention keeps spatial consistency
  • Partial object images and global context are used
  • Scene level data supervises object interactions

Getting started

  • Visit the project page to review the features
  • Download and install the code and dependencies
  • Prepare an input image and run the model

Pros & cons

✓ What we like

  • Generated textures can be refined with MV-Adapter
  • Single object data is used for regularization

! What to watch out for

  • The code library and dependencies must be downloaded and installed
  • An input image is prepared from synthetic, real or stylized data

FAQ

What is the core mechanism?

The card names a multi-instance attention mechanism for object interactions.

Which image types are supported?

The material lists synthetic data, real scene data and stylized scene images.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review