Skip to content

InstructVideo

Guide text-to-video diffusion with reward fine-tuning

Visit official site

What InstructVideo is

The method is described as guiding text-to-video diffusion models through reward fine-tuning with human feedback. It applies reward signals from an established image reward model and is stated to raise the visual quality of generated videos while keeping generalisation.

What you can do with it

  • Fine-tune a text-to-video diffusion model with human feedback
  • Carry out reward fine-tuning through editing
  • Draw reward signals from an image reward model
  • Apply segmented sparse sampling and time-decay rewards

How it works

  • Editing-based fine-tuning is described as reducing cost
  • The same editing approach is described as raising efficiency
  • An image reward model supplies the reward signal
  • Segmented sparse sampling and time-decay rewards shape that signal

Pros & cons

✓ What we like

  • Editing-based fine-tuning is described as reducing cost and raising efficiency
  • The material states strong generalisation is retained

! What to watch out for

  • Reward signals depend on an established image reward model
  • The described scope is text-to-video diffusion models

FAQ

How is the fine-tuning done?

The material describes reward fine-tuning with human feedback carried out through editing.

Where do reward signals come from?

The material names an established image reward model using segmented sparse sampling and time-decay rewards.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review