Iterative Refinement Improves Compositional Image Generation

TL;DR

This work tackles the challenge of faithfully rendering richly compositional prompts in text-to-image generation by introducing an inference-time iterative refinement framework. A T2I generator is guided by a vision-language model critic and verifier, with an image editor applying targeted sub-prompts, all operating under a budget $B = T \times M$ and actions such as STOP, BACKTRACK, RESTART, and CONTINUE to progressively refine outputs. Across ConceptMix, T2I-CompBench, TIIF-Bench, and Visual Jenga, the iterative approach yields consistent gains over compute-matched parallel sampling, with a notable 16.9% improvement in all-correct rate on ConceptMix (k=7) and a full solve-rate upgrade on Visual Jenga from 64.29% to 76.79%; human evaluators also prefer iterative refinements (58.7% vs 41.3%). The method is training-free, model-agnostic, and relies on a simple VLM critic–editor loop, supporting broader applicability of self-correction as a principle for multimodal generation, while also documenting limitations such as potential VLM misreasoning and editor constraints. Formally, the setup decomposes complex prompts into sequential refinements, enabling correct binding of objects, relations, and attributes without bespoke toolchains, and demonstrates robust performance across diverse open- and closed-source T2I models.

Abstract

Text-to-image (T2I) models have achieved remarkable progress, yet they continue to struggle with complex prompts that require simultaneously handling multiple objects, relations, and attributes. Existing inference-time strategies, such as parallel sampling with verifiers or simply increasing denoising steps, can improve prompt alignment but remain inadequate for richly compositional settings where many constraints must be satisfied. Inspired by the success of chain-of-thought reasoning in large language models, we propose an iterative test-time strategy in which a T2I model progressively refines its generations across multiple steps, guided by feedback from a vision-language model as the critic in the loop. Our approach is simple, requires no external tools or priors, and can be flexibly applied to a wide range of image generators and vision-language models. Empirically, we demonstrate consistent gains on image generation across benchmarks: a 16.9% improvement in all-correct rate on ConceptMix (k=7), a 13.8% improvement on T2I-CompBench (3D-Spatial category) and a 12.5% improvement on Visual Jenga scene decomposition compared to compute-matched parallel sampling. Beyond quantitative gains, iterative refinement produces more faithful generations by decomposing complex prompts into sequential corrections, with human evaluators preferring our method 58.7% of the time over 41.3% for the parallel baseline. Together, these findings highlight iterative self-correction as a broadly applicable principle for compositional image generation. Results and visualizations are available at https://iterative-img-gen.github.io/

Figures (16)

• Figure 1: Iterative refinement during inference time enables high fidelity generation of complex prompts on which traditional inference-time scaling strategies such as parallel sampling can fail to generate a fully accurate image even at high num. of samples as shown above.
• Figure 2: Our iterative inference-time strategy achieves strong benefits over computation-matched parallel inference time scaling on multiple state-of-art image generation models.
• Figure 3: Given a complex text prompt $X_p$, a generator $G$ produces an initial image $I_0$. A test-time verifier $V$ and critic $C$, conditioned on $X_p$, output an action--sub-prompt pair $(a_t, p_t)$. The previous image $I_{t-1}$ and sub-prompt $p_t$ are fed to an editor $E$ to yield the next image $I_t$. This process repeats under an inference-time budget $B$, allocated as maximum $T$ iterative rounds over $M$ parallel streams, until a STOP action is emitted or $B$ ($=T\times M$) is exhausted.
• Figure 4: Per-category level improvement for ConceptMix with Qwen-Image. As can be seen the largest improvement for iterative refinement comes from Spatial, Size, Style and Shape categories.
• Figure 5: Comparison against existing test-time scaling methods. As can be seen methods such as GenArtist wang2024genartist and RPG yang2024mastering are difficult to scale due to their reliance on tools or regional priors.