Skip to content

Parrot

A multi-objective RL framework for text-to-image generation

Visit official site

What Parrot is

The framework applies multi-objective reinforcement learning to text-to-image generation and balances rewards through selection.

What you can do with it

  • Select trade-offs between several rewards
  • Jointly optimise a T2I model and a prompt network
  • Apply original-prompt centred guidance at inference

Quality notes

  • Aesthetics, human preference and text-image alignment are named
  • Final image quality is raised by the prompts
  • The output stays faithful to the user input

Pros & cons

✓ What we like

  • Quality-aware prompts raise final image quality
  • Prompt expansion is guarded against forgetting the input

! What to watch out for

  • Trade-offs are selected among several rewards
  • Prompt expansion needs centred guidance at inference

FAQ

What selection method is used?

Batch Pareto optimal selection finds trade-offs between rewards.

Why is guidance added?

It counters forgetting of the original user prompt.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review