POET: Preference Optimization for Enhanced Text-to-Image Generation

Name: POET: Preference Optimization for Enhanced Text-to-Image Generation
Uploaded: 2026-10-03T06:44:30.234367+00:00
Duration: 792 s
Description: This work introduces an input-side scaling framework that uses LLMs to rewrite text prompts before they are fed into a frozen text-to-image model. By training a prompt rewriter using iterative Direct Preference Optimization (DPO) with multiple quality rewards, the method significantly improves image quality, alignment, and aesthetics across various models without retraining the T2I backbone.

This work introduces an input-side scaling framework that uses LLMs to rewrite text prompts before they are fed into a frozen text-to-image model. By training a prompt rewriter using iterative Direct Preference Optimization (DPO) with multiple quality rewards, the method significantly improves image quality, alignment, and aesthetics across various models without retraining the T2I backbone.

Open the full episode