Source-linked AI summary
Compositional Text-to-Image Generation Via Region-aware Bimodal Direct Preference Optimization
Zhuohan Liu, Wujian Peng, Yitong Chen, Zuxuan Wu
TL;DR
Complex prompts remain difficult for text-to-image models to follow. BiDPO combines a bimodal preference dataset, joint text-image preference optimization, and region-level guidance, improving compositional fidelity across benchmarks.
Problem
Text-to-image models still struggle to bind multiple objects and attributes and represent complex inter-object relationships accurately.
Method
BiDPO trains on the BiComp preference dataset, jointly optimizes text and image preferences, and uses region-level guidance for fine-grained alignment.
Results
0.62 vs. 0.53 overall score on GenEval, with BiDPO improving over SDXL across most subtasks and substantially improving compositional generation across four benchmarks.
Takeaways & Limitations
BiDPO demonstrates that preference-based fine-tuning can improve compositional capabilities in text-to-image diffusion models.
Takeaways & Limitations
The data pipeline filters image-caption pairs containing too many objects because detection and segmentation models perform poorly in such cases.
Abstract
from arXiv · showhide
Despite the rapid progress of text-to-image (T2I) models, generating images that accurately reflect complex compositional prompts (covering attribute bindings, object relationships, counting) still remains challenging. To address this, we propose BiDPO, a framework to enhance T2I model's capability of compositional text-to-image generation. We begin by introducing an carefully designed pipeline to construct a large-scale preference dataset, BiComp, with strictly quality control. Then, we extend Diffusion DPO to jointly optimize image and text preferences, which is shown to greatly effective in improving the models to follow complex text prompt in generation. To further enhance the models for fine-grained alignment, we employ a region-level guidance method to focus on regions relevant to compositional concepts. Experimental results demonstrate that our BiDPO substantially improves compositional fidelity, consistently outperforming prior methods across multiple benchmarks. Our approach highlights the potential of preference-based fine-tuning for complex text-to-image tasks, offering a flexible and scalable alternative to existing techniques.
1 Introduction
Existing T2I models generate high-fidelity images but struggle with complex instructions involving multiple objects, attribute bindings, and relationships. BiDPO addresses this through bimodal preference optimization, region-level guidance, and an automated preference-data pipeline for compositional generation.
- Motivation: Existing T2I models struggle with complex prompts involving multiple objects, attribute bindings, and spatial relationships despite strong fidelity and aesthetic quality.
- Method: BiDPO introduces bimodal preference optimization to improve compositional alignment using fine-grained preferences across text and image modalities.
- Dataset: The automated pipeline constructs large-scale, high-quality preference data with textual and visual negative examples for compositional text-to-image training.
- Method: The framework adds region-level guidance that selectively steers model focus toward regions of interest for finer-grained text-to-image alignment.
- Dataset: BiComp contains 57,474 original images and 94,502 edited images spanning color, shape, texture, spatial relationship, non-spatial relationship, and numeracy.
- Experiments: Experiments on widely used benchmarks demonstrate significant performance gains over previous state-of-the-art methods.
2 Related Works
Large-scale diffusion models have advanced text-to-image generation and achieved strong standard-quality performance, yet compositional semantics involving multiple objects, attributes, and relations remain challenging. Preference alignment addresses the gap between generations and human expectations, although image-domain RLHF can be computationally expensive and unstable.
- Text-to-Image Generation: Large-scale diffusion models synthesize realistic images from textual prompts, while compositional semantics involving multiple objects, attributes, and relations remain difficult to capture.Recent systems include Stable Diffusion 3, DALL-E 3, and Flux, which achieve strong performance on standard quality benchmarks.
- Preference Alignment: Preference alignment bridges the gap between model generations and human expectations through reward models, synthetic comparisons, and preference optimization.Early image-domain approaches adapt RLHF and use on-policy algorithms such as PPO.
- Preference Alignment: Image-domain RLHF pipelines are computationally expensive and unstable in high-dimensional image spaces.
3 Method
BiDPO extends diffusion preference optimization to jointly align images and captions for compositional text-to-image generation, while adding region-level guidance for fine-grained alignment. The method is trained on BiComp, a large-scale, high-quality preference dataset with regional annotations constructed through automated caption, detection, segmentation, and editing pipelines.
- Bimodal DPO: BiDPO combines two TextDPO procedures to align image and text preferences, implicitly achieving image-to-image contrastive learning through explicit text-to-text comparisons.The paired losses teach preferences between captions for fixed images and between images for fixed captions.
- Bimodal DPO: TextDPO adapts Diffusion DPO by pairing the preferred image with the less preferred caption, depressing that diffusion process while enhancing the preferred pairing.The model’s noise prediction is conditioned on embeddings for the preferred and less preferred captions.
- Region-level Guidance for Fine-grained Alignment: Region-level guidance weights edited image regions more heavily than uninvolved regions during training to focus the loss on caption-relevant modifications.The guidance mask is generated from bounding boxes obtained in the caption generation and editing pipeline.
- BiComp Data Construction: BiComp addresses the lack of suitable region-annotated composition preference data with a large-scale, high-quality dataset constructed through an automated generation pipeline.The pipeline generates multiple images per caption and creates region annotations for compositional dimensions.
- BiComp Data Construction: The pipeline parses compositional dimensions, detects and segments mentioned objects, describes their attributes or relationships, and synthesizes edited captions and images.It uses DeepSeek models, Grounding Dino, SAM2, Qwen2.5-VL-72B-Instruct, and Qwen-Image-Edit; numeracy captions use direct object counts.
4 Experiments
BiDPO is evaluated across four compositional text-to-image benchmarks and improves over SDXL on complex prompt following, semantic alignment, benchmark robustness, and visual quality. Ablations show that bimodal preference optimization and region-level guidance are important for fine-grained compositional alignment.
- T2I-CompBench: BiDPO improves over SDXL on T2I-CompBench, especially for attribute binding involving color, shape, and texture.The benchmark evaluates object attributes and inter-object relationships.
- GenEval: 0.62 vs. 0.53 overall score: BiDPO improves GenEval performance over SDXL across most complex instruction-following subtasks.GenEval assesses text-to-image models’ ability to follow complex instructions.
- DPG-Bench: BiDPO-SDXL achieves Global 83.92, Entity 85.28, Attribute 85.13, Relation 85.03, Other 84.55, and overall 78.84 on DPG-Bench.The method obtains comparable scores across all reported semantic-alignment categories.
- GenEval 2: GenEval 2 improvements of 6.6% at atomic level and 1.8% at prompt level indicate robustness to benchmark drift.GenEval 2 is described as a more challenging benchmark for modern models.
- Generalization and aesthetic quality: BiDPO improves SD3-Medium as compositional complexity increases, surpasses Flux, and achieves a 2.65% visual-quality improvement on DrawBench.These results support generalization to modern MMDiT models and simultaneous gains in compositionality and aesthetics.
- Ablation studies: Region-level guidance adds 1.2% on T2I-CompBench and 1.4% on GenEval over BiDPO without it, improving fine-grained cross-modal alignment.Ablations also find that supervised fine-tuning is insufficient, image-only DPO helps, and text-only DPO lacks visual guidance.
5 Conclusion
BiDPO introduces bimodal direct preference optimization with region-level guidance for compositional text-to-image generation, trained on the BiComp preference dataset. Experiments on four standard benchmarks show significantly improved compositional capabilities in text-to-image diffusion models.
- BiDPO combines DPO, a bimodal extension, and region-level guidance for compositional text-to-image generation.The method is trained on the composition-aware preference dataset BiComp.
- BiDPO significantly improves the compositional capabilities of text-to-image diffusion models.The improvement is demonstrated through extensive experiments.
- The method is evaluated on four standard benchmarks: T2I-CompBench, GenEval, DPG-Bench, and GenEval 2.
Appendix
The appendix provides supplementary ablation results, data-construction details, and additional visualizations for BiComp and BiDPO.
- Section 6: Section 6 reports the full ablation-study results for SDXL-based models.
- Section 7: Section 7 details the data-construction process, including collected-caption composition and prompts used at different stages.
- Section 8: Section 8 presents additional visualizations of the BiComp dataset and BiDPO method.
6 Ablation Study Details.
The ablation study evaluates SDXL-based models across three compositional text-to-image benchmarks: T2I-CompBench, GenEval, and DPG-Bench.
- Ablation study: The study reports full SDXL-based ablation results in Tables 9, 10, and 11.The cited passage identifies these tables as containing the full ablation results.
- T2I-CompBench: Table 9 presents the ablation study on T2I-CompBench [18].The table is specifically designated for T2I-CompBench.
- GenEval: Table 10 presents the ablation study on GenEval [12].The table is specifically designated for GenEval.
- DPG-Bench: Table 11 presents the ablation study on DPG-Bench [17].The table is specifically designated for DPG-Bench.
7 Data Construction Details
The data-construction pipeline uses LLMs to parse caption dimensions and visible objects, then applies controlled image constraints and dimension-specific VLM prompts to generate and differentiate region information.
- Caption Parsing: DeepSeek-V3 parses each caption into a prioritized dimension, with spatial and non-spatial first, numeracy next, and color, shape, and texture below them.The prompt also assigns “others” the lowest priority.
- Object Parsing: DeepSeek-R1 extracts only concrete, visually visible, detector-detectable objects and excludes adjectives, descriptions, and abstract concepts.The output is organized as an object_list JSON dictionary.
- Image Controls: Images are constrained to one or two objects for color, shape, and texture, exactly two for spatial and non-spatial dimensions, with no repeated object classes.Each object must belong to a unique class.
- Region Information: Qwen2.5-VL-72B-Instruct generates distinct region information using dimension-specific prompts, including dominant colors and spatial relationships.For color, proposed colors must contrast with all provided dominant colors; for spatial cases, the prompt requests a geometrically distinct relationship.
8 More Visualization Results.
This section presents additional visualizations of the BiComp dataset and BiDPO text-to-image generation results. The figures show dataset samples across dimensions and comparisons involving multiple Stable Diffusion variants and IterComp.
- BiComp dataset visualizations: Figures 4 and 5 show BiComp samples for each dimension, pairing images generated from original captions with images generated from edited captions.In each group, the original-caption image is on the left and the edited-caption image is on the right.
- BiDPO generation visualizations: Figure 6 compares Stable Diffusion 3, IterComp, Stable Diffusion XL, and Stable Diffusion XL finetuned with BiDPO.The methods are arranged from left to right in that order.
- BiDPO generation visualizations: Figure 7 compares original Stable Diffusion 1.5 with Stable Diffusion 1.5 finetuned using BiDPO.The visualization focuses on text-to-image generation results for Stable Diffusion 1.5.
- BiDPO generation visualizations: Figure 8 visualizes text-to-image generation results from Stable Diffusion XL.The figure provides additional Stable Diffusion XL generation examples.