Ruoyi Du

Ruoyi Du, Dongliang Chang, Timothy Hospedales et al.

High-resolution image generation with Generative Artificial Intelligence (GenAI) has immense potential but, due to the enormous capital investment required for training, it is increasingly centralised to a few large corporations, and hidden behind paywalls. This paper aims to democratise high-resolution GenAI by advancing the frontier of high-resolution generation while remaining accessible to a broad audience. We demonstrate that existing Latent Diffusion Models (LDMs) possess untapped potential for higher-resolution image generation. Our novel DemoFusion framework seamlessly extends open-source GenAI models, employing Progressive Upscaling, Skip Residual, and Dilated Sampling mechanisms to achieve higher-resolution image generation. The progressive nature of DemoFusion requires more passes, but the intermediate results can serve as "previews", facilitating rapid prompt iteration.

Yiming Zhang, Hong Yu, Ruoyi Du et al.

Audio captioning aims at describing the content of audio clips with human language. Due to the ambiguity of audio, different people may perceive the same audio differently, resulting in caption disparities (i.e., one audio may correlate to several captions with diverse semantics). For that, general audio captioning models achieve the one-to-many training by randomly selecting a correlated caption as the ground truth for each audio. However, it leads to a significant variation in the optimization directions and weakens the model stability. To eliminate this negative effect, in this paper, we propose a two-stage framework for audio captioning: (i) in the first stage, via the contrastive learning, we construct a proxy feature space to reduce the distances between captions correlated to the same audio, and (ii) in the second stage, the proxy feature space is utilized as additional supervision to encourage the model to be optimized in the direction that benefits all the correlated captions. We conducted extensive experiments on two datasets using four commonly used encoder and decoder architectures. Experimental results demonstrate the effectiveness of the proposed method. The code is available at https://github.com/PRIS-CV/Caption-Feature-Space-Regularization.

Tian Zhang, Kongming Liang, Ruoyi Du et al.

Compositional Zero-Shot Learning (CZSL) aims to recognize novel compositions using knowledge learned from seen attribute-object compositions in the training set. Previous works mainly project an image and a composition into a common embedding space to measure their compatibility score. However, both attributes and objects share the visual representations learned above, leading the model to exploit spurious correlations and bias towards seen pairs. Instead, we reconsider CZSL as an out-of-distribution generalization problem. If an object is treated as a domain, we can learn object-invariant features to recognize the attributes attached to any object reliably. Similarly, attribute-invariant features can also be learned when recognizing the objects with attributes as domains. Specifically, we propose an invariant feature learning framework to align different domains at the representation and gradient levels to capture the intrinsic characteristics associated with the tasks. Experiments on two CZSL benchmarks demonstrate that the proposed method significantly outperforms the previous state-of-the-art.

Ruoyi Du, Wenqing Yu, Heqing Wang et al.

As fine-grained visual classification (FGVC) being developed for decades, great works related have exposed a key direction -- finding discriminative local regions and revealing subtle differences. However, unlike identifying visual contents within static images, for recognizing objects in the real physical world, discriminative information is not only present within seen local regions but also hides in other unseen perspectives. In other words, in addition to focusing on the distinguishable part from the whole, for efficient and accurate recognition, it is required to infer the key perspective with a few glances, e.g., people may recognize a "Benz AMG GT" with a glance of its front and then know that taking a look at its exhaust pipe can help to tell which year's model it is. In this paper, back to reality, we put forward the problem of active fine-grained recognition (AFGR) and complete this study in three steps: (i) a hierarchical, multi-view, fine-grained vehicle dataset is collected as the testbed, (ii) a simple experiment is designed to verify that different perspectives contribute differently for FGVC and different categories own different discriminative perspective, (iii) a policy-gradient-based framework is adopted to achieve efficient recognition with active view selection. Comprehensive experiments demonstrate that the proposed method delivers a better performance-efficient trade-off than previous FGVC methods and advanced neural networks.

Qi Qin, Le Zhuo, Yi Xin et al.

We introduce Lumina-Image 2.0, an advanced text-to-image generation framework that achieves significant progress compared to previous work, Lumina-Next. Lumina-Image 2.0 is built upon two key principles: (1) Unification - it adopts a unified architecture (Unified Next-DiT) that treats text and image tokens as a joint sequence, enabling natural cross-modal interactions and allowing seamless task expansion. Besides, since high-quality captioners can provide semantically well-aligned text-image training pairs, we introduce a unified captioning system, Unified Captioner (UniCap), specifically designed for T2I generation tasks. UniCap excels at generating comprehensive and accurate captions, accelerating convergence and enhancing prompt adherence. (2) Efficiency - to improve the efficiency of our proposed model, we develop multi-stage progressive training strategies and introduce inference acceleration techniques without compromising image quality. Extensive evaluations on academic benchmarks and public text-to-image arenas show that Lumina-Image 2.0 delivers strong performances even with only 2.6B parameters, highlighting its scalability and design efficiency. We have released our training details, code, and models at https://github.com/Alpha-VLLM/Lumina-Image-2.0.

Yuqi Yang, Dongliang Chang, Yijia Ling et al.

Colour is one of the most perceptually salient yet least controllable attributes in image generation. Although recent diffusion models can modify object colours from user instructions, their results often deviate from the intended hue, especially for fine-grained and local edits. Early text-driven methods rely on discrete language descriptions that cannot accurately represent continuous chromatic variations. To overcome this limitation, we propose ColourCrafter, a unified diffusion framework that transforms colour editing from global tone transfer into a structured, region-aware generation process. Unlike traditional colour driven methods, ColourCrafter performs token-level fusion of RGB colour tokens and image tokens in latent space, selectively propagating colour information to semantically relevant regions while preserving structural fidelity. A perceptual Lab-space Loss further enhances pixel-level precision by decoupling luminance and chrominance and constraining edits within masked areas. Additionally, we build ColourfulSet, a largescale dataset of high-quality image pairs with continuous and diverse colour variations. Extensive experiments demonstrate that ColourCrafter achieves state-of-the-art colour accuracy, controllability and perceptual fidelity in fine-grained colour editing. Our project is available at https://yangyuqi317.github.io/ColourCrafter.github.io/.

Jiayi Lei, Renrui Zhang, Xiangfei Hu et al.

With the rapid development of diffusion models, text-to-image(T2I) models have made significant progress, showcasing impressive abilities in prompt following and image generation. Recently launched models such as FLUX.1 and Ideogram2.0, along with others like Dall-E3 and Stable Diffusion 3, have demonstrated exceptional performance across various complex tasks, raising questions about whether T2I models are moving towards general-purpose applicability. Beyond traditional image generation, these models exhibit capabilities across a range of fields, including controllable generation, image editing, video, audio, 3D, and motion generation, as well as computer vision tasks like semantic segmentation and depth estimation. However, current evaluation frameworks are insufficient to comprehensively assess these models' performance across expanding domains. To thoroughly evaluate these models, we developed the IMAGINE-E and tested six prominent models: FLUX.1, Ideogram2.0, Midjourney, Dall-E3, Stable Diffusion 3, and Jimeng. Our evaluation is divided into five key domains: structured output generation, realism, and physical consistency, specific domain generation, challenging scenario generation, and multi-style creation tasks. This comprehensive assessment highlights each model's strengths and limitations, particularly the outstanding performance of FLUX.1 and Ideogram2.0 in structured and specific domain tasks, underscoring the expanding applications and potential of T2I models as foundational AI tools. This study provides valuable insights into the current state and future trajectory of T2I models as they evolve towards general-purpose usability. Evaluation scripts will be released at https://github.com/jylei16/Imagine-e.

Z-Image Team, Huanqia Cai, Sihan Cao et al.

The landscape of high-performance image generation models is currently dominated by proprietary systems, such as Nano Banana Pro and Seedream 4.0. Leading open-source alternatives, including Qwen-Image, Hunyuan-Image-3.0 and FLUX.2, are characterized by massive parameter counts (20B to 80B), making them impractical for inference, and fine-tuning on consumer-grade hardware. To address this gap, we propose Z-Image, an efficient 6B-parameter foundation generative model built upon a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture that challenges the "scale-at-all-costs" paradigm. By systematically optimizing the entire model lifecycle -- from a curated data infrastructure to a streamlined training curriculum -- we complete the full training workflow in just 314K H800 GPU hours (approx. $630K). Our few-step distillation scheme with reward post-training further yields Z-Image-Turbo, offering both sub-second inference latency on an enterprise-grade H800 GPU and compatibility with consumer-grade hardware (<16GB VRAM). Additionally, our omni-pre-training paradigm also enables efficient training of Z-Image-Edit, an editing model with impressive instruction-following capabilities. Both qualitative and quantitative experiments demonstrate that our model achieves performance comparable to or surpassing that of leading competitors across various dimensions. Most notably, Z-Image exhibits exceptional capabilities in photorealistic image generation and bilingual text rendering, delivering results that rival top-tier commercial models, thereby demonstrating that state-of-the-art results are achievable with significantly reduced computational overhead. We publicly release our code, weights, and online demo to foster the development of accessible, budget-friendly, yet state-of-the-art generative models.

Dongyang Liu, Peng Gao, David Liu et al.

Diffusion model distillation has emerged as a powerful technique for creating efficient few-step and single-step generators. Among these, Distribution Matching Distillation (DMD) and its variants stand out for their impressive performance, which is widely attributed to their core mechanism of matching the student's output distribution to that of a pre-trained teacher model. In this work, we challenge this conventional understanding. Through a rigorous decomposition of the DMD training objective, we reveal that in complex tasks like text-to-image generation, where CFG is typically required for desirable few-step performance, the primary driver of few-step distillation is not distribution matching, but a previously overlooked component we identify as CFG Augmentation (CA). We demonstrate that this term acts as the core ``engine'' of distillation, while the Distribution Matching (DM) term functions as a ``regularizer'' that ensures training stability and mitigates artifacts. We further validate this decoupling by demonstrating that while the DM term is a highly effective regularizer, it is not unique; simpler non-parametric constraints or GAN-based objectives can serve the same stabilizing function, albeit with different trade-offs. This decoupling of labor motivates a more principled analysis of the properties of both terms, leading to a more systematic and in-depth understanding. This new understanding further enables us to propose principled modifications to the distillation process, such as decoupling the noise schedules for the engine and the regularizer, leading to further performance gains. Notably, our method has been adopted by the Z-Image ( https://github.com/Tongyi-MAI/Z-Image ) project to develop a top-tier 8-step image generation model, empirically validating the generalization and robustness of our findings.

Ruoyi Du, Dongliang Chang, Zhanyu Ma et al.

Despite great strides made on fine-grained visual classification (FGVC), current methods are still heavily reliant on fully-supervised paradigms where ample expert labels are called for. Semi-supervised learning (SSL) techniques, acquiring knowledge from unlabeled data, provide a considerable means forward and have shown great promise for coarse-grained problems. However, exiting SSL paradigms mostly assume in-distribution (i.e., category-aligned) unlabeled data, which hinders their effectiveness when re-proposed on FGVC. In this paper, we put forward a novel design specifically aimed at making out-of-distribution data work for semi-supervised FGVC, i.e., to "clue them in". We work off an important assumption that all fine-grained categories naturally follow a hierarchical structure (e.g., the phylogenetic tree of "Aves" that covers all bird species). It follows that, instead of operating on individual samples, we can instead predict sample relations within this tree structure as the optimization goal of SSL. Beyond this, we further introduced two strategies uniquely brought by these tree structures to achieve inter-sample consistency regularization and reliable pseudo-relation. Our experimental results reveal that (i) the proposed method yields good robustness against out-of-distribution data, and (ii) it can be equipped with prior arts, boosting their performance thus yielding state-of-the-art results. Code is available at https://github.com/PRIS-CV/RelMatch.

Dongliang Chang, Kaiyue Pang, Ruoyi Du et al.

As powerful as fine-grained visual classification (FGVC) is, responding your query with a bird name of "Whip-poor-will" or "Mallard" probably does not make much sense. This however commonly accepted in the literature, underlines a fundamental question interfacing AI and human -- what constitutes transferable knowledge for human to learn from AI? This paper sets out to answer this very question using FGVC as a test bed. Specifically, we envisage a scenario where a trained FGVC model (the AI expert) functions as a knowledge provider in enabling average people (you and me) to become better domain experts ourselves, i.e. those capable in distinguishing between "Whip-poor-will" and "Mallard". Fig. 1 lays out our approach in answering this question. Assuming an AI expert trained using expert human labels, we ask (i) what is the best transferable knowledge we can extract from AI, and (ii) what is the most practical means to measure the gains in expertise given that knowledge? On the former, we propose to represent knowledge as highly discriminative visual regions that are expert-exclusive. For that, we devise a multi-stage learning framework, which starts with modelling visual attention of domain experts and novices before discriminatively distilling their differences to acquire the expert exclusive knowledge. For the latter, we simulate the evaluation process as book guide to best accommodate the learning practice of what is accustomed to humans. A comprehensive human study of 15,000 trials shows our method is able to consistently improve people of divergent bird expertise to recognise once unrecognisable birds. Interestingly, our approach also leads to improved conventional FGVC performance when the extracted knowledge defined is utilised as means to achieve discriminative localisation. Codes are available at: https://github.com/PRIS-CV/Making-a-Bird-AI-Expert-Work-for-You-and-Me

Dongliang Chang, Yixiao Zheng, Zhanyu Ma et al.

Fine-grained visual classification is a challenging task that recognizes the sub-classes belonging to the same meta-class. Large inter-class similarity and intra-class variance is the main challenge of this task. Most exiting methods try to solve this problem by designing complex model structures to explore more minute and discriminative regions. In this paper, we argue that mining multi-regional multi-grained features is precisely the key to this task. Specifically, we introduce a new loss function, termed top-down spatial attention loss (TDSA-Loss), which contains a multi-stage channel constrained module and a top-down spatial attention module. The multi-stage channel constrained module aims to make the feature channels in different stages category-aligned. Meanwhile, the top-down spatial attention module uses the attention map generated by high-level aligned feature channels to make middle-level aligned feature channels to focus on particular regions. Finally, we can obtain multiple discriminative regions on high-level feature channels and obtain multiple more minute regions within these discriminative regions on middle-level feature channels. In summary, we obtain multi-regional multi-grained features. Experimental results over four widely used fine-grained image classification datasets demonstrate the effectiveness of the proposed method. Ablative studies further show the superiority of two modules in the proposed method. Codes are available at: https://github.com/dongliangchang/Top-Down-Spatial-Attention-Loss.

Ruoyi Du, Dongliang Chang, Ayan Kumar Bhunia et al.

Fine-grained visual classification (FGVC) is much more challenging than traditional classification tasks due to the inherently subtle intra-class object variations. Recent works mainly tackle this problem by focusing on how to locate the most discriminative parts, more complementary parts, and parts of various granularities. However, less effort has been placed to which granularities are the most discriminative and how to fuse information cross multi-granularity. In this work, we propose a novel framework for fine-grained visual classification to tackle these problems. In particular, we propose: (i) a progressive training strategy that effectively fuses features from different granularities, and (ii) a random jigsaw patch generator that encourages the network to learn features at specific granularities. We obtain state-of-the-art performances on several standard FGVC benchmark datasets, where the proposed method consistently outperforms existing methods or delivers competitive results. The code will be available at https://github.com/PRIS-CV/PMG-Progressive-Multi-Granularity-Training.

Dengyang Jiang, Xin Jin, Dongyang Liu et al.

The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g, Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for directly continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromises their inherent few-step inference capability. To address this, we propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning. We first find that the modern diffusion model where the LLM/VLM serves as the encoder can inherit its encoder's in-context capabilities. This enables us to make the training as an on-policy self-distillation process. Specifically, during training, we make the model acts as both the teacher and the student with different contexts, where the student is conditioned only on the text feature, while the teacher is conditioned on the multimodal feature of both the text prompt and the target image. Training minimizes the two predicted distributions over the student's own roll-outs. By optimized on the model's own trajectory and under it's own supervision, D-OPSD enables the model to learn new concept, style, etc. without sacrificing the original few-step capacity.

Zhong-Yu Li, Ruoyi Du, Juncheng Yan et al.

Recent progress in diffusion models significantly advances various image generation tasks. However, the current mainstream approach remains focused on building task-specific models, which have limited efficiency when supporting a wide range of different needs. While universal models attempt to address this limitation, they face critical challenges, including generalizable task instruction, appropriate task distributions, and unified architectural design. To tackle these challenges, we propose VisualCloze, a universal image generation framework, which supports a wide range of in-domain tasks, generalization to unseen ones, unseen unification of multiple tasks, and reverse generation. Unlike existing methods that rely on language-based task instruction, leading to task ambiguity and weak generalization, we integrate visual in-context learning, allowing models to identify tasks from visual demonstrations. Meanwhile, the inherent sparsity of visual task distributions hampers the learning of transferable knowledge across tasks. To this end, we introduce Graph200K, a graph-structured dataset that establishes various interrelated tasks, enhancing task density and transferable knowledge. Furthermore, we uncover that our unified image generation formulation shared a consistent objective with image infilling, enabling us to leverage the strong generative priors of pre-trained infilling models without modifying the architectures.

Yupeng Zhou, Zhen Li, Ziheng Ouyang et al.

Encoding videos into discrete tokens could align with text tokens to facilitate concise and unified multi-modal LLMs, yet introducing significant spatiotemporal compression compared to continuous video representation. Previous discrete video VAEs experienced unstable training, long training time, and degraded reconstruction quality. Given the easier training and superior performance of continuous VAEs, an intuitive idea is to enhance discrete video VAEs by leveraging continuous VAEs. After rethinking the intrinsic link between discrete and continuous representations, we found that FSQ could effectively preserve pre-trained continuous VAE priors compared to other quantization methods. By leveraging continuous VAE priors, it converges several times faster than training from scratch and achieves superior performance at convergence. Meanwhile, two structural improvements are proposed. First, inspired by how continuous VAEs enhance reconstruction via enlarged latent dimensions, we introduce a multi-token quantization mechanism, which achieves nearly a 1 dB improvement in PSNR without compromising the token compression ratio. Second, to tackle reconstruction challenges in high-compression video VAEs, we strengthen first-frame reconstruction, enabling the causal VAE to leverage this information in subsequent frames and markedly improving the performance of 4 x 16 x 16 discrete VAEs. Furthermore, we propose a joint discrete-continuous optimization scheme that unifies the two paradigms and, for the first time, achieves competitive performance on both continuous and discrete representations within a single network. We name our method OneVAE to reflect this connection.

Le Zhuo, Ruoyi Du, Han Xiao et al.

Lumina-T2X is a nascent family of Flow-based Large Diffusion Transformers that establishes a unified framework for transforming noise into various modalities, such as images and videos, conditioned on text instructions. Despite its promising capabilities, Lumina-T2X still encounters challenges including training instability, slow inference, and extrapolation artifacts. In this paper, we present Lumina-Next, an improved version of Lumina-T2X, showcasing stronger generation performance with increased training and inference efficiency. We begin with a comprehensive analysis of the Flag-DiT architecture and identify several suboptimal components, which we address by introducing the Next-DiT architecture with 3D RoPE and sandwich normalizations. To enable better resolution extrapolation, we thoroughly compare different context extrapolation methods applied to text-to-image generation with 3D RoPE, and propose Frequency- and Time-Aware Scaled RoPE tailored for diffusion transformers. Additionally, we introduced a sigmoid time discretization schedule to reduce sampling steps in solving the Flow ODE and the Context Drop method to merge redundant visual tokens for faster network evaluation, effectively boosting the overall sampling speed. Thanks to these improvements, Lumina-Next not only improves the quality and efficiency of basic text-to-image generation but also demonstrates superior resolution extrapolation capabilities and multilingual generation using decoder-based LLMs as the text encoder, all in a zero-shot manner. To further validate Lumina-Next as a versatile generative framework, we instantiate it on diverse tasks including visual recognition, multi-view, audio, music, and point cloud generation, showcasing strong performance across these domains. By releasing all codes and model weights, we aim to advance the development of next-generation generative AI capable of universal modeling.

Peng Gao, Le Zhuo, Dongyang Liu et al.

Sora unveils the potential of scaling Diffusion Transformer for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this technical report, we introduce the Lumina-T2X family - a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a unified framework designed to transform noise into images, videos, multi-view 3D objects, and audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as [nextline] and [nextframe] tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. This unified approach enables training within a single framework for different modalities and allows for flexible generation of multimodal data at any resolution, aspect ratio, and length during inference. Advanced techniques like RoPE, RMSNorm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. We expect that the open-sourcing of Lumina-T2X will further foster creativity, transparency, and diversity in the generative AI community.

Jingye Wang, Ruoyi Du, Dongliang Chang et al.

Adaptation to out-of-distribution data is a meta-challenge for all statistical learning algorithms that strongly rely on the i.i.d. assumption. It leads to unavoidable labor costs and confidence crises in realistic applications. For that, domain generalization aims at mining domain-irrelevant knowledge from multiple source domains that can generalize to unseen target domains. In this paper, by leveraging the frequency domain of an image, we uniquely work with two key observations: (i) the high-frequency information of an image depicts object edge structure, which preserves high-level semantic information of the object is naturally consistent across different domains, and (ii) the low-frequency component retains object smooth structure, while this information is susceptible to domain shifts. Motivated by the above observations, we introduce (i) an encoder-decoder structure to disentangle high- and low-frequency feature of an image, (ii) an information interaction mechanism to ensure the helpful knowledge from both two parts can cooperate effectively, and (iii) a novel data augmentation technique that works on the frequency domain to encourage the robustness of frequency-wise feature disentangling. The proposed method obtains state-of-the-art performance on three widely used domain generalization benchmarks (Digit-DG, Office-Home, and PACS).

Siqing Zhang, Ruoyi Du, Dongliang Chang et al.

Fine-grained visual classification (FGVC) aims to distinguish the sub-classes of the same category and its essential solution is to mine the subtle and discriminative regions. Convolution neural networks (CNNs), which employ the cross entropy loss (CE-loss) as the loss function, show poor performance since the model can only learn the most discriminative part and ignore other meaningful regions. Some existing works try to solve this problem by mining more discriminative regions by some detection techniques or attention mechanisms. However, most of them will meet the background noise problem when trying to find more discriminative regions. In this paper, we address it in a knowledge transfer learning manner. Multiple models are trained one by one, and all previously trained models are regarded as teacher models to supervise the training of the current one. Specifically, a orthogonal loss (OR-loss) is proposed to encourage the network to find diverse and meaningful regions. In addition, the first model is trained with only CE-Loss. Finally, all models' outputs with complementary knowledge are combined together for the final prediction result. We demonstrate the superiority of the proposed method and obtain state-of-the-art (SOTA) performances on three popular FGVC datasets.