Mushui Liu

Peng Zhang, Guanghao Zhang, Wanggui He et al.

Recent video multimodal large language models (MLLMs) increasingly couple step-by-step reasoning with on-demand visual evidence retrieval, allowing models to revisit relevant video segments during inference. However, two structural gaps remain in existing thinking-with-video systems. (i) Sampling density is not a learnable decision: existing methods may let the model decide where to look, but the per-window frame rate is largely fixed. As a result, fine-grained evidence is often recovered through repeated retrieval calls, which increases inference context length and training difficulty. (ii) Retrieval and answer generation are usually optimized with a single trajectory-level advantage, so the "where to look" tokens and the "how to answer" tokens receive the same credit even when one is correct and the other is not. To address these gaps, we present DynFrame, a framework that emits the temporal window and the sampling density as native tokens within a single autoregressive pass. This learnable span-density retrieval enables acquiring multi-granularity evidence with a single retrieval step. Based on the above tokenized retrieval interface, we further introduce Segment-Decoupled GRPO (SD-GRPO), which splits each rollout at the retrieval boundary and assigns role-specific token-level advantages, separately crediting the sampling decision and the answer. Trained on the curated DM-CoT-74k and DM-RL-45k, DynFrame-4B is competitive with strong 7B-8B baselines across six benchmarks (NExT-GQA, Charades-STA, ActivityNet-MR, Video-MME, MLVU, LVBench), and DynFrame-8B sets new state-of-the-art on most metrics. Code is available at https://github.com/zhangguanghao523/DynFrame.

Mushui Liu, Bozheng Li, Yunlong Yu

Recent Vision-Language Models (VLMs) \textit{e.g.} CLIP have made great progress in video recognition. Despite the improvement brought by the strong visual backbone in extracting spatial features, CLIP still falls short in capturing and integrating spatial-temporal features which is essential for video recognition. In this paper, we propose OmniCLIP, a framework that adapts CLIP for video recognition by focusing on learning comprehensive features encompassing spatial, temporal, and dynamic spatial-temporal scales, which we refer to as omni-scale features. This is achieved through the design of spatial-temporal blocks that include parallel temporal adapters (PTA), enabling efficient temporal modeling. Additionally, we introduce a self-prompt generator (SPG) module to capture dynamic object spatial features. The synergy between PTA and SPG allows OmniCLIP to discern varying spatial information across frames and assess object scales over time. We have conducted extensive experiments in supervised video recognition, few-shot video recognition, and zero-shot recognition tasks. The results demonstrate the effectiveness of our method, especially with OmniCLIP achieving a top-1 accuracy of 74.30\% on HMDB51 in a 16-shot setting, surpassing the recent MotionPrompt approach even with full training data. The code is available at \url{https://github.com/XiaoBuL/OmniCLIP}.

Jinlong Liu, Wanggui He, Peng Zhang et al.

Reinforcement learning (RL) can improve the prompt following capability of text-to-image (T2I) models, yet obtaining high-quality reward signals remains challenging: CLIP Score is too coarse-grained, while VLM-based reward models (e.g., RewardDance) require costly human-annotated preference data and additional fine-tuning. We propose PromptEcho, a reward construction method that requires \emph{no} annotation and \emph{no} reward model training. Given a generated image and a guiding query, PromptEcho computes the token-level cross-entropy loss of a frozen VLM with the original prompt as the label, directly extracting the image-text alignment knowledge encoded during VLM pretraining. The reward is deterministic, computationally efficient, and improves automatically as stronger open-source VLMs become available. For evaluation, we develop DenseAlignBench, a benchmark of concept-rich dense captions for rigorously testing prompt following capability. Experimental results on two state-of-the-art T2I models (Z-Image and QwenImage-2512) demonstrate that PromptEcho achieves substantial improvements on DenseAlignBench (+26.8pp / +16.2pp net win rate), along with consistent gains on GenEval, DPG-Bench, and TIIFBench without any task-specific training. Ablation studies confirm that PromptEcho comprehensively outperforms inference-based scoring with the same VLM, and that reward quality scales with VLM size. We will open-source the trained models and the DenseAlignBench.

Yunze Tong, Mushui Liu, Canyu Zhao et al.

Deploying GRPO on Flow Matching models has proven effective for text-to-image generation. However, existing paradigms typically propagate an outcome-based reward to all preceding denoising steps without distinguishing the local effect of each step. Moreover, current group-wise ranking mainly compares trajectories at matched timesteps and ignores within-trajectory dependencies, where certain early denoising actions can affect later states via delayed, implicit interactions. We propose TurningPoint-GRPO (TP-GRPO), a GRPO framework that alleviates step-wise reward sparsity and explicitly models long-term effects within the denoising trajectory. TP-GRPO makes two key innovations: (i) it replaces outcome-based rewards with step-level incremental rewards, providing a dense, step-aware learning signal that better isolates each denoising action's "pure" effect, and (ii) it identifies turning points-steps that flip the local reward trend and make subsequent reward evolution consistent with the overall trajectory trend-and assigns these actions an aggregated long-term reward to capture their delayed impact. Turning points are detected solely via sign changes in incremental rewards, making TP-GRPO efficient and hyperparameter-free. Extensive experiments also demonstrate that TP-GRPO exploits reward signals more effectively and consistently improves generation. Demo code is available at https://github.com/YunzeTong/TurningPoint-GRPO.

Fangtai Wu, Hailong Guo, Shijie Huang et al.

Customized image editing aims to equip pre-trained diffusion models with specific visual effects using limited paired data, typically via Low-Rank Adaptation (LoRA). As the number of desired effects grows, storing and dynamically loading numerous these effect LoRAs significantly increases deployment overhead. Furthermore, current pipelines typically cascade these effect LoRAs with acceleration modules for fast generation, which triggers severe parameter interference and results in concept bleeding and style degradation. We propose CollectionLoRA, a multi-teacher on-policy distillation framework capable of distilling the concepts of up to 50 different effect LoRAs along with few-step generation capabilities into a single LoRA. This fundamentally resolves the feature interference issue and significantly reduces deployment costs. Specifically, the method introduces (i) a Probabilistic Dual-Stream Routing mechanism that enables the model to randomly switch between data sources during training, effectively enhancing its generalization in unseen scenarios; (ii) an Asymmetric Orthogonal Prompting strategy to achieve concept isolation within the prompt space; (iii) a Coarse-to-Fine Distillation Objective to mitigate the distribution gap between the teacher and student models. Extensive evaluations show that CollectionLoRA distills all customized effects and few-step generation into a single LoRA, reducing deployment overhead while achieving concept fidelity comparable to or better than independently trained teacher models.

Wanggui He, Siming Fu, Mushui Liu et al.

Auto-regressive models have made significant progress in the realm of language generation, yet they do not perform on par with diffusion models in the domain of image synthesis. In this work, we introduce MARS, a novel framework for T2I generation that incorporates a specially designed Semantic Vision-Language Integration Expert (SemVIE). This innovative component integrates pre-trained LLMs by independently processing linguistic and visual information, freezing the textual component while fine-tuning the visual component. This methodology preserves the NLP capabilities of LLMs while imbuing them with exceptional visual understanding. Building upon the powerful base of the pre-trained Qwen-7B, MARS stands out with its bilingual generative capabilities corresponding to both English and Chinese language prompts and the capacity for joint image and text generation. The flexibility of this framework lends itself to migration towards any-to-any task adaptability. Furthermore, MARS employs a multi-stage training strategy that first establishes robust image-text alignment through complementary bidirectional tasks and subsequently concentrates on refining the T2I generation process, significantly augmenting text-image synchrony and the granularity of image details. Notably, MARS requires only 9% of the GPU days needed by SD1.5, yet it achieves remarkable results across a variety of benchmarks, illustrating the training efficiency and the potential for swift deployment in various applications.

Mushui Liu, Jun Dan, Ziqian Lu et al.

Due to the large-scale image size and object variations, current CNN-based and Transformer-based approaches for remote sensing image semantic segmentation are suboptimal for capturing the long-range dependency or limited to the complex computational complexity. In this paper, we propose CM-UNet, comprising a CNN-based encoder for extracting local image features and a Mamba-based decoder for aggregating and integrating global information, facilitating efficient semantic segmentation of remote sensing images. Specifically, a CSMamba block is introduced to build the core segmentation decoder, which employs channel and spatial attention as the gate activation condition of the vanilla Mamba to enhance the feature interaction and global-local information fusion. Moreover, to further refine the output features from the CNN encoder, a Multi-Scale Attention Aggregation (MSAA) module is employed to merge the different scale features. By integrating the CSMamba block and MSAA module, CM-UNet effectively captures the long-range dependencies and multi-scale global contextual information of large-scale remote-sensing images. Experimental results obtained on three benchmarks indicate that the proposed CM-UNet outperforms existing methods in various performance metrics. The codes are available at https://github.com/XiaoBuL/CM-UNet.

Mushui Liu, Fangtai Wu, Bozheng Li et al.

Few-shot learning (FSL) aims to recognize new concepts using a limited number of visual samples. Existing approaches attempt to incorporate semantic information into the limited visual data for category understanding. However, these methods often enrich class-level feature representations with abstract category names, failing to capture the nuanced features essential for effective generalization. To address this issue, we propose a novel framework for FSL, which incorporates both the abstract class semantics and the concrete class entities extracted from Large Language Models (LLMs), to enhance the representation of the class prototypes. Specifically, our framework composes a Semantic-guided Visual Pattern Extraction (SVPE) module and a Prototype-Calibration (PC) module, where the SVPE meticulously extracts semantic-aware visual patterns across diverse scales, while the PC module seamlessly integrates these patterns to refine the visual prototype, enhancing its representativeness. Extensive experiments on four few-shot classification benchmarks and the BSCD-FSL cross-domain benchmarks showcase remarkable advancements over the current state-of-the-art methods. Notably, for the challenging one-shot setting, our approach, utilizing the ResNet-12 backbone, achieves an impressive average improvement of 1.95% over the second-best competitor.

Bozheng Li, Mushui Liu, Gaoang Wang et al.

In this paper, we propose a novel Temporal Sequence-Aware Model (TSAM) for few-shot action recognition (FSAR), which incorporates a sequential perceiver adapter into the pre-training framework, to integrate both the spatial information and the sequential temporal dynamics into the feature embeddings. Different from the existing fine-tuning approaches that capture temporal information by exploring the relationships among all the frames, our perceiver-based adapter recurrently captures the sequential dynamics alongside the timeline, which could perceive the order change. To obtain the discriminative representations for each class, we extend a textual corpus for each class derived from the large language models (LLMs) and enrich the visual prototypes by integrating the contextual semantic information. Besides, We introduce an unbalanced optimal transport strategy for feature matching that mitigates the impact of class-unrelated features, thereby facilitating more effective decision-making. Experimental results on five FSAR datasets demonstrate that our method set a new benchmark, beating the second-best competitors with large margins.

Mushui Liu, Bozheng Li, Yunlong Yu

Prompt tuning, which involves training a small set of parameters, effectively enhances the pre-trained Vision-Language Models (VLMs) to downstream tasks. However, they often come at the cost of flexibility and adaptability when the tuned models are applied to different datasets or domains. In this paper, we explore capturing the task-specific information via meticulous refinement of entire VLMs, with minimal parameter adjustments. When fine-tuning the entire VLMs for specific tasks under limited supervision, overfitting and catastrophic forgetting become the defacto factors. To mitigate these issues, we propose a framework named CLIP-CITE via designing a discriminative visual-text task, further aligning the visual-text semantics in a supervision manner, and integrating knowledge distillation techniques to preserve the gained knowledge. Extensive experimental results under few-shot learning, base-to-new generalization, domain generalization, and cross-domain generalization settings, demonstrate that our method effectively enhances the performance on specific tasks under limited supervision while preserving the versatility of the VLMs on other datasets.

Weijie He, Mushui Liu, Yunlong Yu

Single-frame infrared small target (SIRST) detection poses a significant challenge due to the requirement to discern minute targets amidst complex infrared background clutter. In this paper, we focus on a weakly-supervised paradigm to obtain high-quality pseudo masks from the point-level annotation by integrating a novel learning-free method with the hybrid of the learning-based method. The learning-free method adheres to a sequential process, progressing from a point annotation to the bounding box that encompasses the target, and subsequently to detailed pseudo masks, while the hybrid is achieved through filtering out false alarms and retrieving missed detections in the network's prediction to provide a reliable supplement for learning-free masks. The experimental results show that our learning-free method generates pseudo masks with an average Intersection over Union (IoU) that is 4.3% higher than the second-best learning-free competitor across three datasets, while the hybrid learning-based method further enhances the quality of pseudo masks, achieving an additional average IoU increase of 3.4%.

Jiacheng Ying, Mushui Liu, Zhe Wu et al.

Blind face restoration has made great progress in producing high-quality and lifelike images. Yet it remains challenging to preserve the ID information especially when the degradation is heavy. Current reference-guided face restoration approaches either require face alignment or personalized test-tuning, which are unfaithful or time-consuming. In this paper, we propose a tuning-free method named RestorerID that incorporates ID preservation during face restoration. RestorerID is a diffusion model-based method that restores low-quality images with varying levels of degradation by using a single reference image. To achieve this, we propose a unified framework to combine the ID injection with the base blind face restoration model. In addition, we design a novel Face ID Rebalancing Adapter (FIR-Adapter) to tackle the problems of content unconsistency and contours misalignment that are caused by information conflicts between the low-quality input and reference image. Furthermore, by employing an Adaptive ID-Scale Adjusting strategy, RestorerID can produce superior restored images across various levels of degradation. Experimental results on the Celeb-Ref dataset and real-world scenarios demonstrate that RestorerID effectively delivers high-quality face restoration with ID preservation, achieving a superior performance compared to the test-tuning approaches and other reference-guided ones. The code of RestorerID is available at \url{https://github.com/YingJiacheng/RestorerID}.

Bozheng Li, Miao Yang, Zhenhan Chen et al. · utoronto

Dynamic Chart Generation (DCG) involves producing code-rendered animated visualizations as charts. While recent advances in multi-modal large language models (MLLMs) have significantly improved their capability on static chart generation and comprehension, MLLMs' potential for handling dynamic chart generation and understanding remains underexplored. To bridge this research gap, we introduce DCG-Bench (Dynamic Chart Generation Benchmark), the first benchmark evaluating MLLM's capability on dynamic chart generation tasks from three dimensions: Simple Text-to-Chart, Detailed Text-to-Chart, and Video-to-Chart tasks. We construct DCG-8K, a high-quality DCG dataset with annotations covering instruction-code-video triplets and QA pairs for both code and video evaluation. Based on DCG-8K, we explored a two-stage training recipe, proposing Joint-Code-Visual Reward for group relative policy optimization to construct expert MLLM Qwen2.5-VL-DCG-3B for the DCG task. Our benchmarking result reveals shortcomings of existing MLLMs in the visual-to-chart task, and our model beats the best open-sourced MLLM with an average 8.31% performance gain across three tasks, and shows on par performance against proprietary models with only 3B parameters, proving the effectiveness of our training recipe. Our code and dataset will be publicly available.

Weijie He, Mushui Liu, Yunlong Yu et al.

Compositional text-to-video generation, which requires synthesizing dynamic scenes with multiple interacting entities and precise spatial-temporal relationships, remains a critical challenge for diffusion-based models. Existing methods struggle with layout discontinuity, entity identity drift, and implausible interaction dynamics due to unconstrained cross-attention mechanisms and inadequate physics-aware reasoning. To address these limitations, we propose DyST-XL, a \textbf{training-free} framework that enhances off-the-shelf text-to-video models (e.g., CogVideoX-5B) through frame-aware control. DyST-XL integrates three key innovations: (1) A Dynamic Layout Planner that leverages large language models (LLMs) to parse input prompts into entity-attribute graphs and generates physics-aware keyframe layouts, with intermediate frames interpolated via trajectory optimization; (2) A Dual-Prompt Controlled Attention Mechanism that enforces localized text-video alignment through frame-aware attention masking, achieving precise control over individual entities; and (3) An Entity-Consistency Constraint strategy that propagates first-frame feature embeddings to subsequent frames during denoising, preserving object identity without manual annotation. Experiments demonstrate that DyST-XL excels in compositional text-to-video generation, significantly improving performance on complex prompts and bridging a crucial gap in training-free video synthesis. The code is released in https://github.com/XiaoBuL/DyST-XL.

Fangtai Wu, Mushui Liu, Weijie He et al.

The unified autoregressive (AR) model excels at multimodal understanding and generation, but its potential for customized image generation remains underexplored. Existing customized generation methods rely on full fine-tuning or adapters, making them costly and prone to overfitting or catastrophic forgetting. In this paper, we propose \textbf{CoAR}, a novel framework for injecting subject concepts into the unified AR models while keeping all pre-trained parameters completely frozen. CoAR learns effective, specific subject representations with only a minimal number of parameters using a Layerwise Multimodal Context Learning strategy. To address overfitting and language drift, we further introduce regularization that preserves the pre-trained distribution and anchors context tokens to improve subject fidelity and re-contextualization. Additionally, CoAR supports training-free subject customization in a user-provided style. Experiments demonstrate that CoAR achieves superior performance on both subject-driven personalization and style personalization, while delivering significant gains in computational and memory efficiency. Notably, CoAR tunes less than \textbf{0.05\%} of the parameters while achieving competitive performance compared to recent Proxy-Tuning. Code: https://github.com/KZF-kzf/CoAR

Yi Wang, Mushui Liu, Wanggui He et al.

Unified generative models have shown remarkable performance in text and image generation. For image synthesis tasks, they adopt straightforward text-to-image (T2I) generation. However, direct T2I generation limits the models in handling complex compositional instructions, which frequently occur in real-world scenarios. Although this issue is vital, existing works mainly focus on improving the basic image generation capability of the models. While such improvements help to some extent, they still fail to adequately resolve the problem. Inspired by Chain of Thought (CoT) solving complex problems step by step, this work aims to introduce CoT into unified generative models to address the challenges of complex image generation that direct T2I generation cannot effectively solve, thereby endowing models with enhanced image generation ability. To achieve this, we first propose Functionality-oriented eXperts (FoXperts), an expert-parallel architecture in our model FoX, which assigns experts by function. FoXperts disentangles potential conflicts in mainstream modality-oriented designs and provides a solid foundation for CoT. When introducing CoT, the first question is how to design it for complex image generation. To this end, we emulate a human-like artistic workflow -- planning, acting, reflection, and correction -- and propose the Multimodal Chain of Thought (MCoT) approach, as the data involves both text and image. To address the subsequent challenge -- designing an effective MCoT training paradigm -- we develop a multi-task joint training scheme that equips the model with all capabilities required for each MCoT step in a disentangled manner. This paradigm avoids the difficulty of collecting consistent multi-step data tuples. Extensive experiments show that FoX consistently outperforms existing unified models on various T2I benchmarks, delivering notable improvements in complex image generation.

Guanghao Zhang, Tao Zhong, Yan Xia et al.

While previous multimodal slow-thinking methods have demonstrated remarkable success in single-image understanding scenarios, their effectiveness becomes fundamentally constrained when extended to more complex multi-image comprehension tasks. This limitation stems from their predominant reliance on text-based intermediate reasoning processes. While for human, when engaging in sophisticated multi-image analysis, they typically perform two complementary cognitive operations: (1) continuous cross-image visual comparison through region-of-interest matching, and (2) dynamic memorization of critical visual concepts throughout the reasoning chain. Motivated by these observations, we propose the Complex Multi-Modal Chain-of-Thought (CMMCoT) framework, a multi-step reasoning framework that mimics human-like "slow thinking" for multi-image understanding. Our approach incorporates two key innovations: 1. The construction of interleaved multimodal multi-step reasoning chains, which utilize critical visual region tokens, extracted from intermediate reasoning steps, as supervisory signals. This mechanism not only facilitates comprehensive cross-modal understanding but also enhances model interpretability. 2. The introduction of a test-time memory augmentation module that expands the model reasoning capacity during inference while preserving parameter efficiency. Furthermore, to facilitate research in this direction, we have curated a novel multi-image slow-thinking dataset. Extensive experiments demonstrate the effectiveness of our model.

D. She, Mushui Liu, Jingxuan Pang et al.

Customized generation has achieved significant progress in image synthesis, yet personalized video generation remains challenging due to temporal inconsistencies and quality degradation. In this paper, we introduce CustomVideoX, an innovative framework leveraging the video diffusion transformer for personalized video generation from a reference image. CustomVideoX capitalizes on pre-trained video networks by exclusively training the LoRA parameters to extract reference features, ensuring both efficiency and adaptability. To facilitate seamless interaction between the reference image and video content, we propose 3D Reference Attention, which enables direct and simultaneous engagement of reference image features with all video frames across spatial and temporal dimensions. To mitigate the excessive influence of reference image features and textual guidance on generated video content during inference, we implement the Time-Aware Reference Attention Bias (TAB) strategy, dynamically modulating reference bias over different time steps. Additionally, we introduce the Entity Region-Aware Enhancement (ERAE) module, aligning highly activated regions of key entity tokens with reference feature injection by adjusting attention bias. To thoroughly evaluate personalized video generation, we establish a new benchmark, VideoBench, comprising over 50 objects and 100 prompts for extensive assessment. Experimental results show that CustomVideoX significantly outperforms existing methods in terms of video consistency and quality.

Zhen Yang, Guibao Shen, Minyang Li et al.

Diffusion models have achieved remarkable progress across various visual generation tasks. However, their performance significantly declines when generating content at resolutions higher than those used during training. Although numerous methods have been proposed to enable high-resolution generation, they all suffer from inefficiency. In this paper, we propose RectifiedHR, a straightforward and efficient solution for training-free high-resolution synthesis. Specifically, we propose a noise refresh strategy that unlocks the model's training-free high-resolution synthesis capability and improves efficiency. Additionally, we are the first to observe the phenomenon of energy decay, which may cause image blurriness during the high-resolution synthesis process. To address this issue, we introduce average latent energy analysis and find that tuning the classifier-free guidance hyperparameter can significantly improve generation performance. Our method is entirely training-free and demonstrates efficient performance. Furthermore, we show that RectifiedHR is compatible with various diffusion model techniques, enabling advanced features such as image editing, customized generation, and video synthesis. Extensive comparisons with numerous baseline methods validate the superior effectiveness and efficiency of RectifiedHR.

Mushui Liu, Weijie He, Ziqian Lu et al.

Prompt learning is a powerful technique for transferring Vision-Language Models (VLMs) such as CLIP to downstream tasks. However, the prompt-based methods that are fine-tuned solely with base classes may struggle to generalize to novel classes in open-vocabulary scenarios, especially when data are limited. To address this issue, we propose an innovative approach called SYNC-CLIP that leverages SYNthetiC data for enhancing the generalization capability of CLIP. Based on the observation of the distribution shift between the real and synthetic samples, we treat real and synthetic samples as distinct domains and propose to optimize separate domain prompts to capture domain-specific information, along with the shared visual prompts to preserve the semantic consistency between two domains. By aligning the cross-domain features, the synthetic data from novel classes can provide implicit guidance to rebalance the decision boundaries. Experimental results on three model generalization tasks demonstrate that our method performs very competitively across various benchmarks. Notably, SYNC-CLIP outperforms the state-of-the-art competitor PromptSRC by an average improvement of 3.0% on novel classes across 11 datasets in open-vocabulary scenarios.

Dong She, Siming Fu, Mushui Liu et al.

Multi-subject personalized generation presents unique challenges in maintaining identity fidelity and semantic coherence when synthesizing images conditioned on multiple reference subjects. Existing methods often suffer from identity blending and attribute leakage due to inadequate modeling of how different subjects should interact within shared representation spaces. We present MOSAIC, a representation-centric framework that rethinks multi-subject generation through explicit semantic correspondence and orthogonal feature disentanglement. Our key insight is that multi-subject generation requires precise semantic alignment at the representation level - knowing exactly which regions in the generated image should attend to which parts of each reference. To enable this, we introduce SemAlign-MS, a meticulously annotated dataset providing fine-grained semantic correspondences between multiple reference subjects and target images, previously unavailable in this domain. Building on this foundation, we propose the semantic correspondence attention loss to enforce precise point-to-point semantic alignment, ensuring high consistency from each reference to its designated regions. Furthermore, we develop the multi-reference disentanglement loss to push different subjects into orthogonal attention subspaces, preventing feature interference while preserving individual identity characteristics. Extensive experiments demonstrate that MOSAIC achieves state-of-the-art performance on multiple benchmarks. Notably, while existing methods typically degrade beyond 3 subjects, MOSAIC maintains high fidelity with 4+ reference subjects, opening new possibilities for complex multi-subject synthesis applications.

Mushui Liu, Yuhang Ma, Yang Zhen et al.

Diffusion models have exhibited substantial success in text-to-image generation. However, they often encounter challenges when dealing with complex and dense prompts involving multiple objects, attribute binding, and long descriptions. In this paper, we propose a novel framework called \textbf{LLM4GEN}, which enhances the semantic understanding of text-to-image diffusion models by leveraging the representation of Large Language Models (LLMs). It can be seamlessly incorporated into various diffusion models as a plug-and-play component. A specially designed Cross-Adapter Module (CAM) integrates the original text features of text-to-image models with LLM features, thereby enhancing text-to-image generation. Additionally, to facilitate and correct entity-attribute relationships in text prompts, we develop an entity-guided regularization loss to further improve generation performance. We also introduce DensePrompts, which contains $7,000$ dense prompts to provide a comprehensive evaluation for the text-to-image generation task. Experiments indicate that LLM4GEN significantly improves the semantic alignment of SD1.5 and SDXL, demonstrating increases of 9.69\% and 12.90\% in color on T2I-CompBench, respectively. Moreover, it surpasses existing models in terms of sample quality, image-text alignment, and human evaluation.