Skip to yearly menu bar Skip to main content


Poster

Self-correcting LLM-controlled Diffusion Models

Tsung-Han Wu · Long Lian · Joseph Gonzalez · Boyi Li · Trevor Darrell

Arch 4A-E Poster #148
[ ] [ Project Page ] [ Paper PDF ]
[ Poster
Wed 19 Jun 5 p.m. PDT — 6:30 p.m. PDT

Abstract:

Text-to-image generation has witnessed significant progress with the advent of diffusion models. Despite the ability to generate photorealistic images, current text-to-image diffusion models still often struggle to accurately interpret and follow complex input text prompts. In contrast to existing models that aim to generate images only with their best effort, we introduce Self-correcting LLM-controlled Diffusion (SLD). SLD is a framework that generates an image from the input prompt, assesses its alignment with the prompt, and performs self-corrections on the inaccuracies in the generated image. Steered by an LLM controller, SLD turns text-to-image generation into an iterative closed-loop process, ensuring correctness in the resulting image. SLD is not only training-free but can also be seamlessly integrated with diffusion models behind API access, such as DALL-E 3, to further boost the performance of state-of-the-art diffusion models. Experimental results show that our approach can rectify a majority of incorrect generations, particularly in generative numeracy, attribute binding, and spatial relationships. Furthermore, by simply adjusting the instructions to the LLM, SLD can perform image editing tasks, bridging the gap between text-to-image generation and image editing pipelines. We will make our code available for future research and applications.

Chat is not available.