LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models (2305.13655v3)

Abstract: Recent advancements in text-to-image diffusion models have yielded impressive results in generating realistic and diverse images. However, these models still struggle with complex prompts, such as those that involve numeracy and spatial reasoning. This work proposes to enhance prompt understanding capabilities in diffusion models. Our method leverages a pretrained LLM for grounded generation in a novel two-stage process. In the first stage, the LLM generates a scene layout that comprises captioned bounding boxes from a given prompt describing the desired image. In the second stage, a novel controller guides an off-the-shelf diffusion model for layout-grounded image generation. Both stages utilize existing pretrained models without additional model parameter optimization. Our method significantly outperforms the base diffusion model and several strong baselines in accurately generating images according to prompts that require various capabilities, doubling the generation accuracy across four tasks on average. Furthermore, our method enables instruction-based multi-round scene specification and can handle prompts in languages not supported by the underlying diffusion model. We anticipate that our method will unleash users' creativity by accurately following more complex prompts. Our code, demo, and benchmark are available at: https://LLM-grounded-diffusion.github.io

Enhancing Text-to-Image Diffusion Models with LLMs

The paper "LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with LLMs" explores an innovative approach to overcoming the limitations inherent in current text-to-image diffusion models, particularly in accurately following complex prompts which involve intricate linguistic constructs such as negation, numeracy, attribute binding, and spatial relationships. The authors propose the integration of LLMs within the diffusion model pipeline to enhance prompt comprehension and thereby improve image generation consistency and precision relative to prompts.

Methodology

The proposed method, termed LLM-grounded Diffusion (LMD), comprises a two-stage generation process:

1. LLM-based Layout Generation:
   • This stage employs an LLM to generate a scene layout from a textual prompt. The layout consists of captioned bounding boxes that delineate object placements in the intended image, alongside background captions and optional negative prompts to omit undesired elements. The LLM leverages in-context learning to produce these layouts by embedding user prompts within predefined templates accompanied by examples.
2. Layout-grounded Image Generation:
   • In the subsequent stage, a layout-grounded controller guides the image generation using the layouts from the LLM. This controller integrates with off-the-shelf latent diffusion models like Stable Diffusion without any additional model training. Using masked latents and cross-attention manipulation, it enables instance-level spatial accuracy in object placement and attribute binding during image synthesis.

Both stages exploit pre-trained models without necessitating further parameter optimization, rendering the approach versatile across various diffusion frameworks without extensive computational overhead.

Results and Impact

The method significantly enhances prompt-following accuracy, with observed improvements of approximately 2.1 to 3.6 times over baseline diffusion models across benchmarks involving complex linguistic constructs such as negation, numeracy, attribute binding, and spatial reasoning. Notably, LMD and its variant LMD+ (integrating pre-trained GLIGEN adapters) exhibited substantial performance gains, particularly in negation and spatial relationship tasks, emphasizing the efficacy of linguistic reasoning provided by LLM integration.

Theoretical and Practical Implications

The introduction of LLMs into the text-to-image generation pipeline presents profound implications both practically and theoretically. Practically, this method empowers users with refined control over image synthesis, accommodating more sophisticated and nuanced linguistic prompts without exhaustive training or fine-tuning of new diffusion models. The theoretical implications suggest a shift towards multi-modal models that leverage cross-domain advancements, where natural language processing can dynamically inform and guide image generation tasks, potentially shaping next-generation models with foundational cross-disciplinary integration.

Future Developments

The framework established by LMD lays ground for future exploration into distilled models that encapsulate this enhanced understanding within unified architectures, potentially bypassing the need for explicit LLM integration during inference stages. Furthermore, fine-tuning open-source LLMs or crafting domain-targeted variants could refine layout generation capabilities, advancing model robustness and alignment with user prompts in varied linguistic contexts.

This research signifies a pivotal step in merging advancements in LLMing with generative image frameworks, unveiling avenues for enriched creative applications and cross-modal innovations in AI research.