Iterative Refinement Improves Compositional Image Generation

Large language models (LLMs) have achieved remarkable progress in recent years, as a result of simply scaling test-time compute. A particularly influential development has been the use of chain-of-thought (CoT) prompting, where models are instructed to "think step by step". Despite its simplicity, this strategy enables models to exhibit sophisticated behaviors such as self-correction, error checking, and iterative refinement.

The success of CoT reasoning in LLMs is closely tied to their pre-training data. During training, LLMs are exposed to large volumes of text that naturally contain traces of human step-by-step reasoning. This supervision on the internet implicitly provides the prior that chain-of-thought prompting later exploits.

By contrast, text-to-image (T2I) models are trained on large-scale datasets of image-caption pairs that lack such structured reasoning traces. As a result, these models do not inherently develop capabilities like self-correction or iterative refinement.

Our central idea is to leverage complementary modules that together mimic the iterative reasoning process observed in LLMs. Concretely, our framework integrates four components: (i) a text-to-image (T2I) model to generate an initial image, (ii) a vision-language model (VLM) critic to identify corrections by comparing the generated image with the target prompt, (iii) an image editor to apply the suggested edits, and (iv) a verifier to evaluate alignment between the final image and the desired description.

We evaluate our approach against the widely adopted strategy of parallel sampling, where multiple images are generated independently and the best one is selected using a verifier. While parallel sampling increases diversity, it does not fundamentally change the underlying generation process.

Our iterative approach consistently outperforms parallel-only baselines across multiple benchmarks, with gains most pronounced on complex compositional tasks.

The gallery contains examples of our method compared to baseline approaches across different types of compositional prompts, and provides the step-by-step reasoning traces in the refinement process.

Text-to-Image Inference-Time Strategies. Recent advances in text-to-image (T2I) generation have demonstrated impressive capabilities. However, complex prompts with multiple objects, relations, and fine-grained attributes remain challenging. Inference-time strategies such as classifier-free guidance, parallel sampling, and grounding-based methods improve prompt fidelity but often fail to scale to richly compositional prompts.

Chain-of-Thought Reasoning in Large Language Models. Chain-of-thought (CoT) prompting has been shown to elicit multi-step reasoning and improve performance on complex language tasks. Drawing inspiration from these strategies, our method applies a similar iterative reasoning paradigm to T2I generation: the critic functions analogously to a CoT process, enabling high-fidelity compositional image synthesis.

Our method is subject to the limitations of its component models, primarily manifesting in two error modes:

• VLM Imperfect Critic: The VLM acting as a critic may fail to accurately identify discrepancies between the image and the text prompt, leading to missing or incorrect feedback.
• Edit Model Imperfect Editor: The image editing model may fail to execute the requested corrections perfectly, sometimes ignoring the edit instruction or introducing artifacts in the process.

We show examples of these failure modes below.