Skip to content

Orthogonal Finetuning (OFT)

Fine-tune text-to-image diffusion models with OFT

Visit official site

What Orthogonal Finetuning (OFT) is

The method is described as orthogonal fine-tuning that guides text-to-image diffusion models toward downstream tasks. It keeps hyperspherical energy between neurons unchanged so the generation ability does not collapse.

What you can do with it

  • Fine-tune DALL-E with few samples for a given subject
  • Generate controllable images from text and control signals
  • Keep generation ability through fine-tuning

Method notes

  • Hyperspherical energy between neurons is held constant
  • Model collapse is prevented
  • Subject-driven and controllable generation are the two tasks named

Pros & cons

✓ What we like

  • Generation ability of the base model is preserved
  • Two fine-tuning tasks are considered in the work

! What to watch out for

  • The approach targets text-to-image diffusion models
  • The described work centres on fine-tuning tasks

FAQ

What does OFT keep unchanged?

The feature list states hyperspherical energy between neurons is kept unchanged.

Which tasks are studied?

The material names subject-driven generation and controllable generation.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review