Recent endeavors in Multimodal Large Language Models (MLLMs) aim to unify visual comprehension and generation by combining LLM and diffusion models, the state-of-the-art in each task, respectively. Existing approaches rely on spatial visual tokens, where image patches are encoded and arranged according to a spatial order (e.g., raster scan). However, we show that spatial tokens lack the recursive structure inherent to languages, hence form an impossible language for LLM to master. In this paper, we build a proper visual language by leveraging diffusion timesteps to learn discrete, recursive visual tokens. Our proposed tokens recursively compensate for the progressive attribute loss in noisy images as timesteps increase, enabling the diffusion model to reconstruct the original image at any timestep. This approach allows us to effectively integrate the strengths of LLMs in autoregressive reasoning and diffusion models in precise image generation, achieving seamless multimodal comprehension and generation within a unified framework. Extensive experiments show that we achieve superior performance for multimodal comprehension and generation simultaneously compared with other MLLMs. Project Page: https://DDT-LLaMA.github.io/.
Guidance techniques are commonly used in diffusion and flow models to improve image quality and consistency for conditional generative tasks such as class-conditional and text-to-image generation. In particular, classifier-free guidance (CFG) -- the most widely adopted guidance technique -- contrasts conditional and unconditional predictions to improve the generated images. This results, however, in trade-offs across quality, diversity and consistency, improving some at the expense of others. While recent work has shown that it is possible to disentangle these factors to some extent, such methods come with an overhead of requiring an additional (weaker) model, or require more forward passes per sampling step. In this paper, we propose Entropy Rectifying Guidance (ERG), a simple and effective guidance mechanism based on inference-time changes in the attention mechanism of state-of-the-art diffusion transformer architectures, which allows for simultaneous improvements over image quality, diversity and prompt consistency. ERG is more general than CFG and similar guidance techniques, as it extends to unconditional sampling. ERG results in significant improvements in various generation tasks such as text-to-image, class-conditional and unconditional image generation. We also show that ERG can be seamlessly combined with other recent guidance methods such as CADS and APG, further boosting generation performance.
We propose a novel framework for ID-preserving generation using a multi-modal encoding strategy rather than injecting identity features via adapters into pre-trained models. Our method treats identity and text as a unified conditioning input. To achieve this, we introduce FaceCLIP, a multi-modal encoder that learns a joint embedding space for both identity and textual semantics. Given a reference face and a text prompt, FaceCLIP produces a unified representation that encodes both identity and text, which conditions a base diffusion model to generate images that are identity-consistent and text-aligned. We also present a multi-modal alignment algorithm to train FaceCLIP, using a loss that aligns its joint representation with face, text, and image embedding spaces. We then build FaceCLIP-SDXL, an ID-preserving image synthesis pipeline by integrating FaceCLIP with Stable Diffusion XL (SDXL). Compared to prior methods, FaceCLIP-SDXL enables photorealistic portrait generation with better identity preservation and textual relevance. Extensive experiments demonstrate its quantitative and qualitative superiority.
Following the initial excitement, Text-to-Image (TTI) models are now being examined more critically. While much of the discourse has focused on biases and stereotypes embedded in large-scale training datasets, the sociotechnical dynamics of user interactions with these models remain underexplored. This study examines the linguistic and semantic choices users make when crafting prompts and how these choices influence the diversity of generated outputs. Analyzing over six million prompts from the Civiverse dataset on the CivitAI platform across seven months, we categorize users into three groups based on their levels of linguistic experimentation: consistent repeaters, occasional repeaters, and non-repeaters. Our findings reveal that as user participation grows over time, prompt language becomes increasingly homogenized through the adoption of popular community tags and descriptors, with repeated prompts comprising 40-50% of submissions. At the same time, semantic similarity and topic preferences remain relatively stable, emphasizing common subjects and surface aesthetics. Using Vendi scores to quantify visual diversity, we demonstrate a clear correlation between lexical similarity in prompts and the visual similarity of generated images, showing that linguistic repetition reinforces less diverse representations. These findings highlight the significant role of user-driven factors in shaping AI-generated imagery, beyond inherent model biases, and underscore the need for tools and practices that encourage greater linguistic and thematic experimentation within TTI systems to foster more inclusive and diverse AI-generated content.
Diffusion models (DMs) have revolutionized text-to-image generation, enabling the creation of highly realistic and customized images from text prompts. With the rise of parameter-efficient fine-tuning (PEFT) techniques like LoRA, users can now customize powerful pre-trained models using minimal computational resources. However, the widespread sharing of fine-tuned DMs on open platforms raises growing ethical and legal concerns, as these models may inadvertently or deliberately generate sensitive or unauthorized content, such as copyrighted material, private individuals, or harmful content. Despite the increasing regulatory attention on generative AI, there are currently no practical tools for systematically auditing these models before deployment. In this paper, we address the problem of concept auditing: determining whether a fine-tuned DM has learned to generate a specific target concept. Existing approaches typically rely on prompt-based input crafting and output-based image classification but suffer from critical limitations, including prompt uncertainty, concept drift, and poor scalability. To overcome these challenges, we introduce Prompt-Agnostic Image-Free Auditing (PAIA), a novel, model-centric concept auditing framework. By treating the DM as the object of inspection, PAIA enables direct analysis of internal model behavior, bypassing the need for optimized prompts or generated images. We evaluate PAIA on 320 controlled model and 690 real-world community models sourced from a public DM sharing platform. PAIA achieves over 90% detection accuracy while reducing auditing time by 18-40x compared to existing baselines. To our knowledge, PAIA is the first scalable and practical solution for pre-deployment concept auditing of diffusion models, providing a practical foundation for safer and more transparent diffusion model sharing.
Quality-Diversity algorithms are powerful tools for discovering diverse, high-performing solutions. Recently, Multi-Objective Quality-Diversity (MOQD) extends QD to problems with several objectives while preserving solution diversity. MOQD has shown promise in fields such as robotics and materials science, where finding trade-offs between competing objectives like energy efficiency and speed, or material properties is essential. However, existing methods in MOQD rely on tessellating the feature space into a grid structure, which prevents their application in domains where feature spaces are unknown or must be learned, such as complex biological systems or latent exploration tasks. In this work, we introduce Multi-Objective Unstructured Repertoire for Quality-Diversity (MOUR-QD), a MOQD algorithm designed for unstructured and unbounded feature spaces. We evaluate MOUR-QD on five robotic tasks. Importantly, we show that our method excels in tasks where features must be learned, paving the way for applying MOQD to unsupervised domains. We also demonstrate that MOUR-QD is advantageous in domains with unbounded feature spaces, outperforming existing grid-based methods. Finally, we demonstrate that MOUR-QD is competitive with established MOQD methods on existing MOQD tasks and achieves double the MOQD-score in some environments. MOUR-QD opens up new opportunities for MOQD in domains like protein design and image generation.
Blood vessel networks in the brain play a crucial role in stroke research, where understanding their topology is essential for analyzing blood flow dynamics. However, extracting detailed topological vessel network information from microscopy data remains a significant challenge, mainly due to the scarcity of labeled training data and the need for high topological accuracy. This work combines synthetic data generation with deep learning to automatically extract vessel networks as graphs from volumetric microscopy data. To combat data scarcity, we introduce a comprehensive pipeline for generating large-scale synthetic datasets that mirror the characteristics of real vessel networks. Our three-stage approach progresses from abstract graph generation through vessel mask creation to realistic medical image synthesis, incorporating biological constraints and imaging artifacts at each stage. Using this synthetic data, we develop a two-stage deep learning pipeline of 3D U-Net-based models for node detection and edge prediction. Fine-tuning on real microscopy data shows promising adaptation, improving edge prediction F1 scores from 0.496 to 0.626 by training on merely 5 manually labeled samples. These results suggest that automated vessel network extraction is becoming practically feasible, opening new possibilities for large-scale vascular analysis in stroke research.
Large text-to-image diffusion models have demonstrated remarkable image synthesis capabilities, but their indiscriminate training on Internet-scale data has led to learned concepts that enable harmful, copyrighted, or otherwise undesirable content generation. We address the task of concept erasure in diffusion models, i.e., removing a specified concept from a pre-trained model such that prompting the concept (or related synonyms) no longer yields its depiction, while preserving the model's ability to generate other content. We propose a novel method, Attentional Concept Erasure (ACE), that integrates a closed-form attention manipulation with lightweight fine-tuning. Theoretically, we formulate concept erasure as aligning the model's conditional distribution on the target concept with a neutral distribution. Our approach identifies and nullifies concept-specific latent directions in the cross-attention modules via a gated low-rank adaptation, followed by adversarially augmented fine-tuning to ensure thorough erasure of the concept and its synonyms. Empirically, we demonstrate on multiple benchmarks, including object classes, celebrity faces, explicit content, and artistic styles, that ACE achieves state-of-the-art concept removal efficacy and robustness. Compared to prior methods, ACE better balances generality (erasing concept and related terms) and specificity (preserving unrelated content), scales to dozens of concepts, and is efficient, requiring only a few seconds of adaptation per concept. We will release our code to facilitate safer deployment of diffusion models.
Diffusion Transformer (DiT), an emerging diffusion model for visual generation, has demonstrated superior performance but suffers from substantial computational costs. Our investigations reveal that these costs primarily stem from the \emph{static} inference paradigm, which inevitably introduces redundant computation in certain \emph{diffusion timesteps} and \emph{spatial regions}. To overcome this inefficiency, we propose \textbf{Dy}namic \textbf{Di}ffusion \textbf{T}ransformer (DyDiT), an architecture that \emph{dynamically} adjusts its computation along both \emph{timestep} and \emph{spatial} dimensions. Specifically, we introduce a \emph{Timestep-wise Dynamic Width} (TDW) approach that adapts model width conditioned on the generation timesteps. In addition, we design a \emph{Spatial-wise Dynamic Token} (SDT) strategy to avoid redundant computation at unnecessary spatial locations. TDW and SDT can be seamlessly integrated into DiT and significantly accelerates the generation process. Building on these designs, we further enhance DyDiT in three key aspects. First, DyDiT is integrated seamlessly with flow matching-based generation, enhancing its versatility. Furthermore, we enhance DyDiT to tackle more complex visual generation tasks, including video generation and text-to-image generation, thereby broadening its real-world applications. Finally, to address the high cost of full fine-tuning and democratize technology access, we investigate the feasibility of training DyDiT in a parameter-efficient manner and introduce timestep-based dynamic LoRA (TD-LoRA). Extensive experiments on diverse visual generation models, including DiT, SiT, Latte, and FLUX, demonstrate the effectiveness of DyDiT.
Variational Autoencoders (VAEs) are powerful generative models capable of learning compact latent representations. However, conventional VAEs often generate relatively blurry images due to their assumption of an isotropic Gaussian latent space and constraints in capturing high-frequency details. In this paper, we explore a novel wavelet-based approach (Wavelet-VAE) in which the latent space is constructed using multi-scale Haar wavelet coefficients. We propose a comprehensive method to encode the image features into multi-scale detail and approximation coefficients and introduce a learnable noise parameter to maintain stochasticity. We thoroughly discuss how to reformulate the reparameterization trick, address the KL divergence term, and integrate wavelet sparsity principles into the training objective. Our experimental evaluation on CIFAR-10 and other high-resolution datasets demonstrates that the Wavelet-VAE improves visual fidelity and recovers higher-resolution details compared to conventional VAEs. We conclude with a discussion of advantages, potential limitations, and future research directions for wavelet-based generative modeling.
Deep learning based rendering has demonstrated major improvements for photo-realistic image synthesis, applicable to various applications including visual effects in movies and photo-realistic scene building in video games. However, a significant limitation is the difficulty of decomposing the illumination and material parameters, which limits such methods to reconstruct an input scene, without any possibility to control these parameters. This paper introduces a novel physics based neural deferred shading pipeline to decompose the data-driven rendering process, learn a generalizable shading function to produce photo-realistic results for shading and relighting tasks, we also provide a shadow estimator to efficiently mimic shadowing effect. Our model achieves improved performance compared to classical models and a state-of-art neural shading model, and enables generalizable photo-realistic shading from arbitrary illumination input.
The comic production industry requires reference-based line art colorization with high accuracy, efficiency, contextual consistency, and flexible control. A comic page often involves diverse characters, objects, and backgrounds, which complicates the coloring process. Despite advancements in diffusion models for image generation, their application in line art colorization remains limited, facing challenges related to handling extensive reference images, time-consuming inference, and flexible control. We investigate the necessity of extensive contextual image guidance on the quality of line art colorization. To address these challenges, we introduce Cobra, an efficient and versatile method that supports color hints and utilizes over 200 reference images while maintaining low latency. Central to Cobra is a Causal Sparse DiT architecture, which leverages specially designed positional encodings, causal sparse attention, and Key-Value Cache to effectively manage long-context references and ensure color identity consistency. Results demonstrate that Cobra achieves accurate line art colorization through extensive contextual reference, significantly enhancing inference speed and interactivity, thereby meeting critical industrial demands. We release our codes and models on our project page: https://zhuang2002.github.io/Cobra/.
With the rapid advancement of text-to-image (T2I) generation models, assessing the semantic alignment between generated images and text descriptions has become a significant research challenge. Current methods, including those based on Visual Question Answering (VQA), still struggle with fine-grained assessments and precise quantification of image-text alignment. This paper presents an improved evaluation method named Instruction-augmented Multimodal Alignment for Image-Text and Element Matching (iMatch), which evaluates image-text semantic alignment by fine-tuning multimodal large language models. We introduce four innovative augmentation strategies: First, the QAlign strategy creates a precise probabilistic mapping to convert discrete scores from multimodal large language models into continuous matching scores. Second, a validation set augmentation strategy uses pseudo-labels from model predictions to expand training data, boosting the model's generalization performance. Third, an element augmentation strategy integrates element category labels to refine the model's understanding of image-text matching. Fourth, an image augmentation strategy employs techniques like random lighting to increase the model's robustness. Additionally, we propose prompt type augmentation and score perturbation strategies to further enhance the accuracy of element assessments. Our experimental results show that the iMatch method significantly surpasses existing methods, confirming its effectiveness and practical value. Furthermore, our iMatch won first place in the CVPR NTIRE 2025 Text to Image Generation Model Quality Assessment - Track 1 Image-Text Alignment.
Advancements in image generation technologies have raised significant concerns about their potential misuse, such as producing misinformation and deepfakes. Therefore, there is an urgent need for effective methods to detect AI-generated images (AIGI). Despite progress in AIGI detection, achieving reliable performance across diverse generation models and scenes remains challenging due to the lack of source-invariant features and limited generalization capabilities in existing methods. In this work, we explore the potential of using image entropy as a cue for AIGI detection and propose Multi-granularity Local Entropy Patterns (MLEP), a set of entropy feature maps computed across shuffled small patches over multiple image scaled. MLEP comprehensively captures pixel relationships across dimensions and scales while significantly disrupting image semantics, reducing potential content bias. Leveraging MLEP, a robust CNN-based classifier for AIGI detection can be trained. Extensive experiments conducted in an open-world scenario, evaluating images synthesized by 32 distinct generative models, demonstrate significant improvements over state-of-the-art methods in both accuracy and generalization.
In this work, we present SupResDiffGAN, a novel hybrid architecture that combines the strengths of Generative Adversarial Networks (GANs) and diffusion models for super-resolution tasks. By leveraging latent space representations and reducing the number of diffusion steps, SupResDiffGAN achieves significantly faster inference times than other diffusion-based super-resolution models while maintaining competitive perceptual quality. To prevent discriminator overfitting, we propose adaptive noise corruption, ensuring a stable balance between the generator and the discriminator during training. Extensive experiments on benchmark datasets show that our approach outperforms traditional diffusion models such as SR3 and I$^2$SB in efficiency and image quality. This work bridges the performance gap between diffusion- and GAN-based methods, laying the foundation for real-time applications of diffusion models in high-resolution image generation.
Diffusion models have become the most popular approach for high-quality image generation, but their high computational cost still remains a significant challenge. To address this problem, we propose U-Shape Mamba (USM), a novel diffusion model that leverages Mamba-based layers within a U-Net-like hierarchical structure. By progressively reducing sequence length in the encoder and restoring it in the decoder through Mamba blocks, USM significantly lowers computational overhead while maintaining strong generative capabilities. Experimental results against Zigma, which is currently the most efficient Mamba-based diffusion model, demonstrate that USM achieves one-third the GFlops, requires less memory and is faster, while outperforming Zigma in image quality. Frechet Inception Distance (FID) is improved by 15.3, 0.84 and 2.7 points on AFHQ, CelebAHQ and COCO datasets, respectively. These findings highlight USM as a highly efficient and scalable solution for diffusion-based generative models, making high-quality image synthesis more accessible to the research community while reducing computational costs.
Despite recent advances in diffusion models, achieving reliable image generation and editing remains challenging due to the inherent diversity induced by stochastic noise in the sampling process. Instruction-guided image editing with diffusion models offers user-friendly capabilities, yet editing failures, such as background distortion, frequently occur. Users often resort to trial and error, adjusting seeds or prompts to achieve satisfactory results, which is inefficient. While seed selection methods exist for Text-to-Image (T2I) generation, they depend on external verifiers, limiting applicability, and evaluating multiple seeds increases computational complexity. To address this, we first establish a multiple-seed-based image editing baseline using background consistency scores, achieving Best-of-N performance without supervision. Building on this, we introduce ELECT (Early-timestep Latent Evaluation for Candidate Selection), a zero-shot framework that selects reliable seeds by estimating background mismatches at early diffusion timesteps, identifying the seed that retains the background while modifying only the foreground. ELECT ranks seed candidates by a background inconsistency score, filtering unsuitable samples early based on background consistency while preserving editability. Beyond standalone seed selection, ELECT integrates into instruction-guided editing pipelines and extends to Multimodal Large-Language Models (MLLMs) for joint seed and prompt selection, further improving results when seed selection alone is insufficient. Experiments show that ELECT reduces computational costs (by 41 percent on average and up to 61 percent) while improving background consistency and instruction adherence, achieving around 40 percent success rates in previously failed cases - without any external supervision or training.
State-of-the-art visual generative AI tools hold immense potential to assist users in the early ideation stages of creative tasks -- offering the ability to generate (rather than search for) novel and unprecedented (instead of existing) images of considerable quality that also adhere to boundless combinations of user specifications. However, many large-scale text-to-image systems are designed for broad applicability, yielding conventional output that may limit creative exploration. They also employ interaction methods that may be difficult for beginners. Given that creative end users often operate in diverse, context-specific ways that are often unpredictable, more variation and personalization are necessary. We introduce POET, a real-time interactive tool that (1) automatically discovers dimensions of homogeneity in text-to-image generative models, (2) expands these dimensions to diversify the output space of generated images, and (3) learns from user feedback to personalize expansions. An evaluation with 28 users spanning four creative task domains demonstrated POET's ability to generate results with higher perceived diversity and help users reach satisfaction in fewer prompts during creative tasks, thereby prompting them to deliberate and reflect more on a wider range of possible produced results during the co-creative process. Focusing on visual creativity, POET offers a first glimpse of how interaction techniques of future text-to-image generation tools may support and align with more pluralistic values and the needs of end users during the ideation stages of their work.
Generating high-quality, photorealistic textures for 3D human avatars remains a fundamental yet challenging task in computer vision and multimedia field. However, real paired front and back images of human subjects are rarely available with privacy, ethical and cost of acquisition, which restricts scalability of the data. Additionally, learning priors from image inputs using deep generative models, such as GANs or diffusion models, to infer unseen regions such as the human back often leads to artifacts, structural inconsistencies, or loss of fine-grained detail. To address these issues, we present SMPL-GPTexture (skinned multi-person linear model - general purpose Texture), a novel pipeline that takes natural language prompts as input and leverages a state-of-the-art text-to-image generation model to produce paired high-resolution front and back images of a human subject as the starting point for texture estimation. Using the generated paired dual-view images, we first employ a human mesh recovery model to obtain a robust 2D-to-3D SMPL alignment between image pixels and the 3D model's UV coordinates for each views. Second, we use an inverted rasterization technique that explicitly projects the observed colour from the input images into the UV space, thereby producing accurate, complete texture maps. Finally, we apply a diffusion-based inpainting module to fill in the missing regions, and the fusion mechanism then combines these results into a unified full texture map. Extensive experiments shows that our SMPL-GPTexture can generate high resolution texture aligned with user's prompts.
After demonstrating significant success in image synthesis, Generative Adversarial Network (GAN) models have likewise made significant progress in the field of speech synthesis, leveraging their capacity to adapt the precise distribution of target data through adversarial learning processes. Notably, in the realm of State-Of-The-Art (SOTA) GAN-based Voice Conversion (VC) models, there exists a substantial disparity in naturalness between real and GAN-generated speech samples. Furthermore, while many GAN models currently operate on a single generator discriminator learning approach, optimizing target data distribution is more effectively achievable through a single generator multi-discriminator learning scheme. Hence, this study introduces a novel GAN model named Collective Learning Mechanism-based Optimal Transport GAN (CLOT-GAN) model, incorporating multiple discriminators, including the Deep Convolutional Neural Network (DCNN) model, Vision Transformer (ViT), and conformer. The objective of integrating various discriminators lies in their ability to comprehend the formant distribution of mel-spectrograms, facilitated by a collective learning mechanism. Simultaneously, the inclusion of Optimal Transport (OT) loss aims to precisely bridge the gap between the source and target data distribution, employing the principles of OT theory. The experimental validation on VCC 2018, VCTK, and CMU-Arctic datasets confirms that the CLOT-GAN-VC model outperforms existing VC models in objective and subjective assessments.