Junshi Huang

Hu Xu, Zhaolong Xing, Congcong Liu et al.

Post-training pruning compresses large language models to high sparsity using a small unlabelled calibration set, and recent work has concluded that the choice of calibration source has only modest impact on averaged post-pruning accuracy. We ask whether this conclusion survives once calibration impact is evaluated separately across distinct capability dimensions rather than aggregated. Decomposing post-pruning capability into General, Commonsense, Code, and Math, and analysing $n{=}15$ calibration sources via Spearman correlations between OIT information metrics and per-dimension retention, we uncover an opposite-sign trade-off: calibration perplexity correlates positively with General retention ($ρ{=}{+}0.71$) but negatively with Math and Code retention ($ρ{=}{-}0.53,\,{-}0.59$; $p{<}0.05$), so no single source can preserve all capabilities. We respond with multi-source calibration mixing, and propose IGSP, an information-guided self-calibration protocol that automates multi-source construction without capability-aligned corpora by minimising 4-gram aggregation and balancing perplexity across dimensions. On LLaMA-3.1-8B at SparseGPT 60% sparsity, a uniform multi-source mix reaches 58.8% total retention, outperforming the best single source (MetaMath, 50.0%) by $+8.8$ and the C4 default (40.0%) by $+18.8$; IGSP improves over Self-Cal by $+2.4$ and SGS by $+4.8$.

Zhengcong Fei, Mingyuan Fan, Changqian Yu et al.

In this paper, we present DiT-MoE, a sparse version of the diffusion Transformer, that is scalable and competitive with dense networks while exhibiting highly optimized inference. The DiT-MoE includes two simple designs: shared expert routing and expert-level balance loss, thereby capturing common knowledge and reducing redundancy among the different routed experts. When applied to conditional image generation, a deep analysis of experts specialization gains some interesting observations: (i) Expert selection shows preference with spatial position and denoising time step, while insensitive with different class-conditional information; (ii) As the MoE layers go deeper, the selection of experts gradually shifts from specific spacial position to dispersion and balance. (iii) Expert specialization tends to be more concentrated at the early time step and then gradually uniform after half. We attribute it to the diffusion process that first models the low-frequency spatial information and then high-frequency complex information. Based on the above guidance, a series of DiT-MoE experimentally achieves performance on par with dense networks yet requires much less computational load during inference. More encouragingly, we demonstrate the potential of DiT-MoE with synthesized image data, scaling diffusion model at a 16.5B parameter that attains a new SoTA FID-50K score of 1.80 in 512$\times$512 resolution settings. The project page: https://github.com/feizc/DiT-MoE.

Zhengcong Fei, Junshi Huang, Xiaoming Wei et al.

Existing approaches to image captioning usually generate the sentence word-by-word from left to right, with the constraint of conditioned on local context including the given image and history generated words. There have been many studies target to make use of global information during decoding, e.g., iterative refinement. However, it is still under-explored how to effectively and efficiently incorporate the future context. To respond to this issue, inspired by that Non-Autoregressive Image Captioning (NAIC) can leverage two-side relation with modified mask operation, we aim to graft this advance to the conventional Autoregressive Image Captioning (AIC) model while maintaining the inference efficiency without extra time cost. Specifically, AIC and NAIC models are first trained combined with shared visual encoders, forcing the visual encoder to contain sufficient and valid future context; then the AIC model is encouraged to capture the causal dynamics of cross-layer interchanging from NAIC model on its unconfident words, which follows a teacher-student paradigm and optimized with the distribution calibration training objective. Empirical evidences demonstrate that our proposed approach clearly surpass the state-of-the-art baselines in both automatic metrics and human evaluations on the MS COCO benchmark. The source code is available at: https://github.com/feizc/Future-Caption.

Zhengcong Fei, Mingyuan Fan, Changqian Yu et al.

This paper explores a simple extension of diffusion-based rectified flow Transformers for text-to-music generation, termed as FluxMusic. Generally, along with design in advanced Flux\footnote{https://github.com/black-forest-labs/flux} model, we transfers it into a latent VAE space of mel-spectrum. It involves first applying a sequence of independent attention to the double text-music stream, followed by a stacked single music stream for denoised patch prediction. We employ multiple pre-trained text encoders to sufficiently capture caption semantic information as well as inference flexibility. In between, coarse textual information, in conjunction with time step embeddings, is utilized in a modulation mechanism, while fine-grained textual details are concatenated with the music patch sequence as inputs. Through an in-depth study, we demonstrate that rectified flow training with an optimized architecture significantly outperforms established diffusion methods for the text-to-music task, as evidenced by various automatic metrics and human preference evaluations. Our experimental data, code, and model weights are made publicly available at: \url{https://github.com/feizc/FluxMusic}.

Duojun Huang, Jichang Li, Weikai Chen et al.

Active domain adaptation (ADA) aims to improve the model adaptation performance by incorporating active learning (AL) techniques to label a maximally-informative subset of target samples. Conventional AL methods do not consider the existence of domain shift, and hence, fail to identify the truly valuable samples in the context of domain adaptation. To accommodate active learning and domain adaption, the two naturally different tasks, in a collaborative framework, we advocate that a customized learning strategy for the target data is the key to the success of ADA solutions. We present Divide-and-Adapt (DiaNA), a new ADA framework that partitions the target instances into four categories with stratified transferable properties. With a novel data subdivision protocol based on uncertainty and domainness, DiaNA can accurately recognize the most gainful samples. While sending the informative instances for annotation, DiaNA employs tailored learning strategies for the remaining categories. Furthermore, we propose an informativeness score that unifies the data partitioning criteria. This enables the use of a Gaussian mixture model (GMM) to automatically sample unlabeled data into the proposed four categories. Thanks to the "divideand-adapt" spirit, DiaNA can handle data with large variations of domain gap. In addition, we show that DiaNA can generalize to different domain adaptation settings, such as unsupervised domain adaptation (UDA), semi-supervised domain adaptation (SSDA), source-free domain adaptation (SFDA), etc.

Kaiheng Weng, Xiangxiang Chu, Xiaoming Xu et al.

We present a hardware-efficient architecture of convolutional neural network, which has a repvgg-like architecture. Flops or parameters are traditional metrics to evaluate the efficiency of networks which are not sensitive to hardware including computing ability and memory bandwidth. Thus, how to design a neural network to efficiently use the computing ability and memory bandwidth of hardware is a critical problem. This paper proposes a method how to design hardware-aware neural network. Based on this method, we designed EfficientRep series convolutional networks, which are high-computation hardware(e.g. GPU) friendly and applied in YOLOv6 object detection framework. YOLOv6 has published YOLOv6N/YOLOv6S/YOLOv6M/YOLOv6L models in v1 and v2 versions.

Zihan Ding, Tianrui Hui, Junshi Huang et al.

Referring video object segmentation aims to predict foreground labels for objects referred by natural language expressions in videos. Previous methods either depend on 3D ConvNets or incorporate additional 2D ConvNets as encoders to extract mixed spatial-temporal features. However, these methods suffer from spatial misalignment or false distractors due to delayed and implicit spatial-temporal interaction occurring in the decoding phase. To tackle these limitations, we propose a Language-Bridged Duplex Transfer (LBDT) module which utilizes language as an intermediary bridge to accomplish explicit and adaptive spatial-temporal interaction earlier in the encoding phase. Concretely, cross-modal attention is performed among the temporal encoder, referring words and the spatial encoder to aggregate and transfer language-relevant motion and appearance information. In addition, we also propose a Bilateral Channel Activation (BCA) module in the decoding phase for further denoising and highlighting the spatial-temporal consistent features via channel-wise activation. Extensive experiments show our method achieves new state-of-the-art performances on four popular benchmarks with 6.8% and 6.9% absolute AP gains on A2D Sentences and J-HMDB Sentences respectively, while consuming around 7x less computational overhead.

Zhengcong Fei, Mingyuan Fan, Junshi Huang

Recent works on personalized text-to-image generation usually learn to bind a special token with specific subjects or styles of a few given images by tuning its embedding through gradient descent. It is natural to question whether we can optimize the textual inversions by only accessing the process of model inference. As only requiring the forward computation to determine the textual inversion retains the benefits of less GPU memory, simple deployment, and secure access for scalable models. In this paper, we introduce a \emph{gradient-free} framework to optimize the continuous textual inversion in an iterative evolutionary strategy. Specifically, we first initialize an appropriate token embedding for textual inversion with the consideration of visual and text vocabulary information. Then, we decompose the optimization of evolutionary strategy into dimension reduction of searching space and non-convex gradient-free optimization in subspace, which significantly accelerates the optimization process with negligible performance loss. Experiments in several applications demonstrate that the performance of text-to-image model equipped with our proposed gradient-free method is comparable to that of gradient-based counterparts with variant GPU/CPU platforms, flexible employment, as well as computational efficiency.

Zhengcong Fei, Mingyuan Fan, Li Zhu et al.

It is well believed that the higher uncertainty in a word of the caption, the more inter-correlated context information is required to determine it. However, current image captioning methods usually consider the generation of all words in a sentence sequentially and equally. In this paper, we propose an uncertainty-aware image captioning framework, which parallelly and iteratively operates insertion of discontinuous candidate words between existing words from easy to difficult until converged. We hypothesize that high-uncertainty words in a sentence need more prior information to make a correct decision and should be produced at a later stage. The resulting non-autoregressive hierarchy makes the caption generation explainable and intuitive. Specifically, we utilize an image-conditioned bag-of-word model to measure the word uncertainty and apply a dynamic programming algorithm to construct the training pairs. During inference, we devise an uncertainty-adaptive parallel beam search technique that yields an empirically logarithmic time complexity. Extensive experiments on the MS COCO benchmark reveal that our approach outperforms the strong baseline and related methods on both captioning quality as well as decoding speed.

Zhengcong Fei, Mingyuan Fan, Junshi Huang

This paper presents that the masked-modeling principle driving the success of large foundational vision models can be effectively applied to audio by making predictions in a latent space. We introduce Audio-based Joint-Embedding Predictive Architecture (A-JEPA), a simple extension method for self-supervised learning from the audio spectrum. Following the design of I-JEPA, our A-JEPA encodes visible audio spectrogram patches with a curriculum masking strategy via context encoder, and predicts the representations of regions sampled at well-designed locations. The target representations of those regions are extracted by the exponential moving average of context encoder, \emph{i.e.}, target encoder, on the whole spectrogram. We find it beneficial to transfer random block masking into time-frequency aware masking in a curriculum manner, considering the complexity of highly correlated in local time and frequency in audio spectrograms. To enhance contextual semantic understanding and robustness, we fine-tune the encoder with a regularized masking on target datasets, instead of input dropping or zero. Empirically, when built with Vision Transformers structure, we find A-JEPA to be highly scalable and sets new state-of-the-art performance on multiple audio and speech classification tasks, outperforming other recent models that use externally supervised pre-training.

Zihan Ding, Zi-han Ding, Tianrui Hui et al.

Panoptic Narrative Grounding (PNG) is an emerging task whose goal is to segment visual objects of things and stuff categories described by dense narrative captions of a still image. The previous two-stage approach first extracts segmentation region proposals by an off-the-shelf panoptic segmentation model, then conducts coarse region-phrase matching to ground the candidate regions for each noun phrase. However, the two-stage pipeline usually suffers from the performance limitation of low-quality proposals in the first stage and the loss of spatial details caused by region feature pooling, as well as complicated strategies designed for things and stuff categories separately. To alleviate these drawbacks, we propose a one-stage end-to-end Pixel-Phrase Matching Network (PPMN), which directly matches each phrase to its corresponding pixels instead of region proposals and outputs panoptic segmentation by simple combination. Thus, our model can exploit sufficient and finer cross-modal semantic correspondence from the supervision of densely annotated pixel-phrase pairs rather than sparse region-phrase pairs. In addition, we also propose a Language-Compatible Pixel Aggregation (LCPA) module to further enhance the discriminative ability of phrase features through multi-round refinement, which selects the most compatible pixels for each phrase to adaptively aggregate the corresponding visual context. Extensive experiments show that our method achieves new state-of-the-art performance on the PNG benchmark with 4.0 absolute Average Recall gains.

Tianrui Hui, Zihan Ding, Junshi Huang et al.

Panoptic narrative grounding (PNG) aims to segment things and stuff objects in an image described by noun phrases of a narrative caption. As a multimodal task, an essential aspect of PNG is the visual-linguistic interaction between image and caption. The previous two-stage method aggregates visual contexts from offline-generated mask proposals to phrase features, which tend to be noisy and fragmentary. The recent one-stage method aggregates only pixel contexts from image features to phrase features, which may incur semantic misalignment due to lacking object priors. To realize more comprehensive visual-linguistic interaction, we propose to enrich phrases with coupled pixel and object contexts by designing a Phrase-Pixel-Object Transformer Decoder (PPO-TD), where both fine-grained part details and coarse-grained entity clues are aggregated to phrase features. In addition, we also propose a PhraseObject Contrastive Loss (POCL) to pull closer the matched phrase-object pairs and push away unmatched ones for aggregating more precise object contexts from more phrase-relevant object tokens. Extensive experiments on the PNG benchmark show our method achieves new state-of-the-art performance with large margins.

Zhengcong Fei, Mingyuan Fan, Li Zhu et al.

Recently, Vector Quantized AutoRegressive (VQ-AR) models have shown remarkable results in text-to-image synthesis by equally predicting discrete image tokens from the top left to bottom right in the latent space. Although the simple generative process surprisingly works well, is this the best way to generate the image? For instance, human creation is more inclined to the outline-to-fine of an image, while VQ-AR models themselves do not consider any relative importance of image patches. In this paper, we present a progressive model for high-fidelity text-to-image generation. The proposed method takes effect by creating new image tokens from coarse to fine based on the existing context in a parallel manner, and this procedure is recursively applied with the proposed error revision mechanism until an image sequence is completed. The resulting coarse-to-fine hierarchy makes the image generation process intuitive and interpretable. Extensive experiments in MS COCO benchmark demonstrate that the progressive model produces significantly better results compared with the previous VQ-AR method in FID score across a wide variety of categories and aspects. Moreover, the design of parallel generation in each step allows more than $\times 13$ inference acceleration with slight performance loss.

Zhengcong Fei, Shuman Tian, Junshi Huang et al.

Ensemble of machine learning models yields improved performance as well as robustness. However, their memory requirements and inference costs can be prohibitively high. Knowledge distillation is an approach that allows a single model to efficiently capture the approximate performance of an ensemble while showing poor scalability as demand for re-training when introducing new teacher models. In this paper, we study if we can utilize the meta-learning strategy to directly predict the parameters of a single model with comparable performance of an ensemble. Hereto, we introduce WeightFormer, a Transformer-based model that can predict student network weights layer by layer in a forward pass, according to the teacher model parameters. The proprieties of WeightFormer are investigated on the CIFAR-10, CIFAR-100, and ImageNet datasets for model structures of VGGNet-11, ResNet-50, and ViT-B/32, where it demonstrates that our method can achieve approximate classification performance of an ensemble and outperforms both the single network and standard knowledge distillation. More encouragingly, we show that WeightFormer results can further exceeds average ensemble with minor fine-tuning. Importantly, our task along with the model and results can potentially lead to a new, more efficient, and scalable paradigm of ensemble networks parameter learning.

Yuchen Ma, Zhengcong Fei, Junshi Huang

Recently, the tokens of images share the same static data flow in many dense networks. However, challenges arise from the variance among the objects in images, such as large variations in the spatial scale and difficulties of recognition for visual entities. In this paper, we propose a data-dependent token routing strategy to elaborate the routing paths of image tokens for Dynamic Vision Transformer, dubbed DiT. The proposed framework generates a data-dependent path per token, adapting to the object scales and visual discrimination of tokens. In feed-forward, the differentiable routing gates are designed to select the scaling paths and feature transformation paths for image tokens, leading to multi-path feature propagation. In this way, the impact of object scales and visual discrimination of image representation can be carefully tuned. Moreover, the computational cost can be further reduced by giving budget constraints to the routing gate and early-stopping of feature extraction. In experiments, our DiT achieves superior performance and favorable complexity/accuracy trade-offs than many SoTA methods on ImageNet classification, object detection, instance segmentation, and semantic segmentation. Particularly, the DiT-B5 obtains 84.8\% top-1 Acc on ImageNet with 10.3 GFLOPs, which is 1.0\% higher than that of the SoTA method with similar computational complexity. These extensive results demonstrate that DiT can serve as versatile backbones for various vision tasks.

Zhengcong Fei, Mingyuan Fan, Changqian Yu et al.

This paper presents a new exploration into a category of diffusion models built upon state space architecture. We endeavor to train diffusion models for image data, wherein the traditional U-Net backbone is supplanted by a state space backbone, functioning on raw patches or latent space. Given its notable efficacy in accommodating long-range dependencies, Diffusion State Space Models (DiS) are distinguished by treating all inputs including time, condition, and noisy image patches as tokens. Our assessment of DiS encompasses both unconditional and class-conditional image generation scenarios, revealing that DiS exhibits comparable, if not superior, performance to CNN-based or Transformer-based U-Net architectures of commensurate size. Furthermore, we analyze the scalability of DiS, gauged by the forward pass complexity quantified in Gflops. DiS models with higher Gflops, achieved through augmentation of depth/width or augmentation of input tokens, consistently demonstrate lower FID. In addition to demonstrating commendable scalability characteristics, DiS-H/2 models in latent space achieve performance levels akin to prior diffusion models on class-conditional ImageNet benchmarks at the resolution of 256$\times$256 and 512$\times$512, while significantly reducing the computational burden. The code and models are available at: https://github.com/feizc/DiS.

Di Qiu, Zheng Chen, Rui Wang et al.

Recent advancements in character video synthesis still depend on extensive fine-tuning or complex 3D modeling processes, which can restrict accessibility and hinder real-time applicability. To address these challenges, we propose a simple yet effective tuning-free framework for character video synthesis, named MovieCharacter, designed to streamline the synthesis process while ensuring high-quality outcomes. Our framework decomposes the synthesis task into distinct, manageable modules: character segmentation and tracking, video object removal, character motion imitation, and video composition. This modular design not only facilitates flexible customization but also ensures that each component operates collaboratively to effectively meet user needs. By leveraging existing open-source models and integrating well-established techniques, MovieCharacter achieves impressive synthesis results without necessitating substantial resources or proprietary datasets. Experimental results demonstrate that our framework enhances the efficiency, accessibility, and adaptability of character video synthesis, paving the way for broader creative and interactive applications.

Lichen Ma, Zipeng Guo, Yu He et al.

To circumvent the inherent fidelity bottlenecks and optimization misalignment of VAE-based latent diffusion, pixel-space diffusion models have emerged as a compelling end-to-end paradigm. However, existing pixel diffusion models often struggle to balance computational efficiency with the preservation of high-frequency details. They frequently resort to patch-based compression or restricted local decoding, leading to a "spectral compromise" where high-frequency and fine-grained pixel information are suppressed. To address these challenges, we propose \textbf{FrequencyBooster}, a novel framework designed to empower pixel diffusion with full-frequency modeling capabilities without prohibitive overhead. The core of our method is a high-capacity decoder that specializes in extracting exhaustive high-frequency details and low-frequency semantics, the latter of which is derived from a Diffusion Transformer (DiT) backbone. Unlike prior works that sacrifice global context for local refinement, FrequencyBooster leverages high-dimensional feature representations to maintain global structural integrity while achieving superior pixel-level precision. Extensive experiments on ImageNet demonstrate the effectiveness of our approach: our model achieves a state-of-the-art FID of \textbf{1.60} at $256 \times 256$ resolution within only 320 epochs. Furthermore, at $512 \times 512$ resolution, FrequencyBooster attains an FID of \textbf{1.69}, significantly outperforming existing pixel-space and latent-space generative models.

Yu He, Lichen Ma, Zipeng Guo et al.

Pixel-space diffusion models bypass the reconstruction bottleneck of Variational Autoencoders (VAEs) but face a fundamental "granularity dilemma": capturing global semantics favors large patch scales, while generating high-fidelity details demands fine-grained inputs. To address this issue, we propose HyperDiT, a unified framework establishing Hyper-Connected Cross-Scale Interactions to bridge the semantic and pixel manifold. Diverging from injecting semantics by AdaLN, HyperDiT utilizes Cross-Attention mechanisms, enabling fine-grained tokens to query multi-level semantic anchors globally. To resolve the spatial mismatch during multi-scale interactions, we introduce Scale-Aware Rotary Position Embedding (SA-RoPE) to ensure precise geometric alignment among tokens of varying patch sizes. Furthermore, we incorporate Registers to learn the dense semantics from a pretrained Visual Foundation Model (VFM), effectively reducing generation hallucination and artifacts. Extensive experiments demonstrate that HyperDiT achieves state-of-the-art (SoTA) FID of $\mathbf{1.56}$ on ImageNet $256\times256$ directly within the pixel space. By combining the fine-grained stream with semantic guidance, HyperDiT offers a superior paradigm for high-fidelity pixel generation.

Qili Wang, Dajiang Wu, Zihang Xu et al.

Significant progress has been made in talking-face video generation research; however, precise lip-audio synchronization and high visual quality remain challenging in editing lip shapes based on input audio. This paper introduces JoyGen, a novel two-stage framework for talking-face generation, comprising audio-driven lip motion generation and visual appearance synthesis. In the first stage, a 3D reconstruction model and an audio2motion model predict identity and expression coefficients respectively. Next, by integrating audio features with a facial depth map, we provide comprehensive supervision for precise lip-audio synchronization in facial generation. Additionally, we constructed a Chinese talking-face dataset containing 130 hours of high-quality video. JoyGen is trained on the open-source HDTF dataset and our curated dataset. Experimental results demonstrate superior lip-audio synchronization and visual quality achieved by our method.

Yu He, Fang Li, Haoyang Tong et al.

Recent advances in generative models have empowered impressive layered image generation, yet their success is largely confined to graphic design domains. The layering of in-the-wild images remains an underexplored problem, limiting fine-grained editing and applications of images in real-world scenarios. Specifically, challenges remain in scalable layered data and the modeling of object interaction in natural images, such as illumination effects and structural boundary. To address these bottlenecks, we propose a novel framework for high-fidelity natural image decomposition. First, we introduce an Agent-driven Data Decomposition (ADD) pipeline that orchestrates agents and tools to synthesize layered data without manual intervention. Utilizing this pipeline, we construct a large-scale dataset, named LiWi-100k, with over 100,000 high-quality layered in-the-wild images. Second, we present a novel framework that jointly improves photometric fidelity and alpha boundary accuracy. Specifically, shadow-guided learning explicitly models the illumination effects, and degradation-restoration objective provides boundary-correction supervision by recovering clean foreground image from degraded one. Extensive experiments demonstrate that our framework achieves state-of-the-art (SoTA) performance in natural image decomposition, outperforming existing models in RGB L1 and Alpha IoU metrics. We will soon release our code and dataset.

Yu He, Ting Zhu, Yichun Liu et al.

Recent research work on fashion outfit generation focuses on promoting visual consistency of garments by leveraging key information from reference image and text prompt. However, the potential of outfit generation remains underexplored, requiring comprehensive e-commercial dataset and elaborative utilization of multi-modal condition. In this paper, we propose a brand-new e-commerce dataset, named Fashion130k, with various occasions, models, and garment types. For the consistent generation of garment, we design a framework with Unified Multi-modal Condition (UMC) to align and integrate the text and visual prompts into generation model. Specifically, we explore an embedding refiner to extract the unified embeddings of multi-modal prompts, within which a Fusion Transformer is proposed to align the multi-modal embeddings by adjusting the modality gap between text and image. Based on unified embeddings, the attention in generation model is redesigned to emphasis the correlations between prompts and noise image, inducing that the noise image can select the pivotal tokens of prompts for consistent outfit generation. Our dataset and proposed framework offer a general and nuanced exploration of multi-modal prompts for generation models. Extensive experiments on real-world applications and benchmark demonstrate the effectiveness of UMC in visual consistency, achieving promising result than that of SoTA methods.

Zipeng Guo, Lichen Ma, Xiaolong Fu et al.

In web data, product images are central to boosting user engagement and advertising efficacy on e-commerce platforms, yet the intrusive elements such as watermarks and promotional text remain major obstacles to delivering clear and appealing product visuals. Although diffusion-based inpainting methods have advanced, they still face challenges in commercial settings due to unreliable object removal and limited domain-specific adaptation. To tackle these challenges, we propose Repainter, a reinforcement learning framework that integrates spatial-matting trajectory refinement with Group Relative Policy Optimization (GRPO). Our approach modulates attention mechanisms to emphasize background context, generating higher-reward samples and reducing unwanted object insertion. We also introduce a composite reward mechanism that balances global, local, and semantic constraints, effectively reducing visual artifacts and reward hacking. Additionally, we contribute EcomPaint-100K, a high-quality, large-scale e-commerce inpainting dataset, and a standardized benchmark EcomPaint-Bench for fair evaluation. Extensive experiments demonstrate that Repainter significantly outperforms state-of-the-art methods, especially in challenging scenes with intricate compositions. We will release our code and weights upon acceptance.

Lit Sin Tan, Junzhe Chen, Xiaolong Fu et al.

Existing test-time scaling (TTS) methods for unified multimodal models (UMMs) in text-to-image (T2I) generation primarily rely on search or sampling strategies that produce only instance-level improvements, limiting the ability to learn from prior inferences and accumulate knowledge across similar prompts. To overcome these limitations, we propose Meta-TTRL, a metacognitive test-time reinforcement learning framework. Meta-TTRL performs test-time parameter optimization guided by model-intrinsic monitoring signals derived from the meta-knowledge of UMMs, achieving self-improvement and capability-level improvement at test time. Extensive experiments demonstrate that Meta-TTRL generalizes well across three representative UMMs, including Janus-Pro-7B, BAGEL, and Qwen-Image, achieving significant gains on compositional reasoning tasks and multiple T2I benchmarks with limited data. We provide the first comprehensive analysis to investigate the potential of test-time reinforcement learning (TTRL) for T2I generation in UMMs. Our analysis further reveals a key insight underlying effective TTRL: metacognitive synergy, where monitoring signals align with the model's optimization regime to enable self-improvement.

Zhengcong Fei, Mingyuan Fan, Changqian Yu et al.

Transformers have catalyzed advancements in computer vision and natural language processing (NLP) fields. However, substantial computational complexity poses limitations for their application in long-context tasks, such as high-resolution image generation. This paper introduces a series of architectures adapted from the RWKV model used in the NLP, with requisite modifications tailored for diffusion model applied to image generation tasks, referred to as Diffusion-RWKV. Similar to the diffusion with Transformers, our model is designed to efficiently handle patchnified inputs in a sequence with extra conditions, while also scaling up effectively, accommodating both large-scale parameters and extensive datasets. Its distinctive advantage manifests in its reduced spatial aggregation complexity, rendering it exceptionally adept at processing high-resolution images, thereby eliminating the necessity for windowing or group cached operations. Experimental results on both condition and unconditional image generation tasks demonstrate that Diffison-RWKV achieves performance on par with or surpasses existing CNN or Transformer-based diffusion models in FID and IS metrics while significantly reducing total computation FLOP usage.

Lichen Ma, Xiaolong Fu, Gaojing Zhou et al.

With the rapid advancement of image generation, visual text editing using natural language instructions has received increasing attention. The main challenge of this task is to fully understand the instruction and reference image, and thus generate visual text that is style-consistent with the image. Previous methods often involve complex steps of specifying the text content and attributes, such as font size, color, and layout, without considering the stylistic consistency with the reference image. To address this, we propose UM-Text, a unified multimodal model for context understanding and visual text editing by natural language instructions. Specifically, we introduce a Visual Language Model (VLM) to process the instruction and reference image, so that the text content and layout can be elaborately designed according to the context information. To generate an accurate and harmonious visual text image, we further propose the UM-Encoder to combine the embeddings of various condition information, where the combination is automatically configured by VLM according to the input instruction. During training, we propose a regional consistency loss to offer more effective supervision for glyph generation on both latent and RGB space, and design a tailored three-stage training strategy to further enhance model performance. In addition, we contribute the UM-DATA-200K, a large-scale visual text image dataset on diverse scenes for model training. Extensive qualitative and quantitative results on multiple public benchmarks demonstrate that our method achieves state-of-the-art performance.

Xiaoyue Duan, Shuhao Cui, Guoliang Kang et al.

Consistent editing of real images is a challenging task, as it requires performing non-rigid edits (e.g., changing postures) to the main objects in the input image without changing their identity or attributes. To guarantee consistent attributes, some existing methods fine-tune the entire model or the textual embedding for structural consistency, but they are time-consuming and fail to perform non-rigid edits. Other works are tuning-free, but their performances are weakened by the quality of Denoising Diffusion Implicit Model (DDIM) reconstruction, which often fails in real-world scenarios. In this paper, we present a novel approach called Tuning-free Inversion-enhanced Control (TIC), which directly correlates features from the inversion process with those from the sampling process to mitigate the inconsistency in DDIM reconstruction. Specifically, our method effectively obtains inversion features from the key and value features in the self-attention layers, and enhances the sampling process by these inversion features, thus achieving accurate reconstruction and content-consistent editing. To extend the applicability of our method to general editing scenarios, we also propose a mask-guided attention concatenation strategy that combines contents from both the inversion and the naive DDIM editing processes. Experiments show that the proposed method outperforms previous works in reconstruction and consistent editing, and produces impressive results in various settings.

Zhengcong Fei, Mingyuan Fan, Junshi Huang

Consistency models have exhibited remarkable capabilities in facilitating efficient image/video generation, enabling synthesis with minimal sampling steps. It has proven to be advantageous in mitigating the computational burdens associated with diffusion models. Nevertheless, the application of consistency models in music generation remains largely unexplored. To address this gap, we present Music Consistency Models (\texttt{MusicCM}), which leverages the concept of consistency models to efficiently synthesize mel-spectrogram for music clips, maintaining high quality while minimizing the number of sampling steps. Building upon existing text-to-music diffusion models, the \texttt{MusicCM} model incorporates consistency distillation and adversarial discriminator training. Moreover, we find it beneficial to generate extended coherent music by incorporating multiple diffusion processes with shared constraints. Experimental results reveal the effectiveness of our model in terms of computational efficiency, fidelity, and naturalness. Notable, \texttt{MusicCM} achieves seamless music synthesis with a mere four sampling steps, e.g., only one second per minute of the music clip, showcasing the potential for real-time application.

Xiaolong Fu, Lichen Ma, Zipeng Guo et al.

The integration of Reinforcement Learning (RL) into flow matching models for text-to-image (T2I) generation has driven substantial advances in generation quality. However, these gains often come at the cost of exhaustive exploration and inefficient sampling strategies due to slight variation in the sampling group. Building on this insight, we propose Dynamic-TreeRPO, which implements the sliding-window sampling strategy as a tree-structured search with dynamic noise intensities along depth. We perform GRPO-guided optimization and constrained Stochastic Differential Equation (SDE) sampling within this tree structure. By sharing prefix paths of the tree, our design effectively amortizes the computational overhead of trajectory search. With well-designed noise intensities for each tree layer, Dynamic-TreeRPO can enhance the variation of exploration without any extra computational cost. Furthermore, we seamlessly integrate Supervised Fine-Tuning (SFT) and RL paradigm within Dynamic-TreeRPO to construct our proposed LayerTuning-RL, reformulating the loss function of SFT as a dynamically weighted Progress Reward Model (PRM) rather than a separate pretraining method. By associating this weighted PRM with dynamic-adaptive clipping bounds, the disruption of exploration process in Dynamic-TreeRPO is avoided. Benefiting from the tree-structured sampling and the LayerTuning-RL paradigm, our model dynamically explores a diverse search space along effective directions. Compared to existing baselines, our approach demonstrates significant superiority in terms of semantic consistency, visual fidelity, and human preference alignment on established benchmarks, including HPS-v2.1, PickScore, and ImageReward. In particular, our model outperforms SoTA by $4.9\%$, $5.91\%$, and $8.66\%$ on those benchmarks, respectively, while improving the training efficiency by nearly $50\%$.

Zhengcong Fei, Mingyuan Fan, Changqian Yu et al.

This paper unveils Dimba, a new text-to-image diffusion model that employs a distinctive hybrid architecture combining Transformer and Mamba elements. Specifically, Dimba sequentially stacked blocks alternate between Transformer and Mamba layers, and integrate conditional information through the cross-attention layer, thus capitalizing on the advantages of both architectural paradigms. We investigate several optimization strategies, including quality tuning, resolution adaption, and identify critical configurations necessary for large-scale image generation. The model's flexible design supports scenarios that cater to specific resource constraints and objectives. When scaled appropriately, Dimba offers substantial throughput and a reduced memory footprint relative to conventional pure Transformers-based benchmarks. Extensive experiments indicate that Dimba achieves comparable performance compared with benchmarks in terms of image quality, artistic rendering, and semantic control. We also report several intriguing properties of architecture discovered during evaluation and release checkpoints in experiments. Our findings emphasize the promise of large-scale hybrid Transformer-Mamba architectures in the foundational stage of diffusion models, suggesting a bright future for text-to-image generation.

Mingyuan Fan, Shenqi Lai, Junshi Huang et al.

BiSeNet has been proved to be a popular two-stream network for real-time segmentation. However, its principle of adding an extra path to encode spatial information is time-consuming, and the backbones borrowed from pretrained tasks, e.g., image classification, may be inefficient for image segmentation due to the deficiency of task-specific design. To handle these problems, we propose a novel and efficient structure named Short-Term Dense Concatenate network (STDC network) by removing structure redundancy. Specifically, we gradually reduce the dimension of feature maps and use the aggregation of them for image representation, which forms the basic module of STDC network. In the decoder, we propose a Detail Aggregation module by integrating the learning of spatial information into low-level layers in single-stream manner. Finally, the low-level features and deep features are fused to predict the final segmentation results. Extensive experiments on Cityscapes and CamVid dataset demonstrate the effectiveness of our method by achieving promising trade-off between segmentation accuracy and inference speed. On Cityscapes, we achieve 71.9% mIoU on the test set with a speed of 250.4 FPS on NVIDIA GTX 1080Ti, which is 45.2% faster than the latest methods, and achieve 76.8% mIoU with 97.0 FPS while inferring on higher resolution images.

Xuanyi Dong, Junshi Huang, Yi Yang et al.

In this paper, we present a novel and general network structure towards accelerating the inference process of convolutional neural networks, which is more complicated in network structure yet with less inference complexity. The core idea is to equip each original convolutional layer with another low-cost collaborative layer (LCCL), and the element-wise multiplication of the ReLU outputs of these two parallel layers produces the layer-wise output. The combined layer is potentially more discriminative than the original convolutional layer, and its inference is faster for two reasons: 1) the zero cells of the LCCL feature maps will remain zero after element-wise multiplication, and thus it is safe to skip the calculation of the corresponding high-cost convolution in the original convolutional layer, 2) LCCL is very fast if it is implemented as a 1*1 convolution or only a single filter shared by all channels. Extensive experiments on the CIFAR-10, CIFAR-100 and ILSCRC-2012 benchmarks show that our proposed network structure can accelerate the inference process by 32\% on average with negligible performance drop.

Junshi Huang, Rogerio S. Feris, Qiang Chen et al.

We address the problem of cross-domain image retrieval, considering the following practical application: given a user photo depicting a clothing image, our goal is to retrieve the same or attribute-similar clothing items from online shopping stores. This is a challenging problem due to the large discrepancy between online shopping images, usually taken in ideal lighting/pose/background conditions, and user photos captured in uncontrolled conditions. To address this problem, we propose a Dual Attribute-aware Ranking Network (DARN) for retrieval feature learning. More specifically, DARN consists of two sub-networks, one for each domain, whose retrieval feature representations are driven by semantic attribute learning. We show that this attribute-guided learning is a key factor for retrieval accuracy improvement. In addition, to further align with the nature of the retrieval problem, we impose a triplet visual similarity constraint for learning to rank across the two sub-networks. Another contribution of our work is a large-scale dataset which makes the network learning feasible. We exploit customer review websites to crawl a large set of online shopping images and corresponding offline user photos with fine-grained clothing attributes, i.e., around 450,000 online shopping images and about 90,000 exact offline counterpart images of those online ones. All these images are collected from real-world consumer websites reflecting the diversity of the data modality, which makes this dataset unique and rare in the academic community. We extensively evaluate the retrieval performance of networks in different configurations. The top-20 retrieval accuracy is doubled when using the proposed DARN other than the current popular solution using pre-trained CNN features only (0.570 vs. 0.268).

Yunchao Wei, Wei Xia, Junshi Huang et al.

Convolutional Neural Network (CNN) has demonstrated promising performance in single-label image classification tasks. However, how CNN best copes with multi-label images still remains an open problem, mainly due to the complex underlying object layouts and insufficient multi-label training images. In this work, we propose a flexible deep CNN infrastructure, called Hypotheses-CNN-Pooling (HCP), where an arbitrary number of object segment hypotheses are taken as the inputs, then a shared CNN is connected with each hypothesis, and finally the CNN output results from different hypotheses are aggregated with max pooling to produce the ultimate multi-label predictions. Some unique characteristics of this flexible deep CNN infrastructure include: 1) no ground truth bounding box information is required for training; 2) the whole HCP infrastructure is robust to possibly noisy and/or redundant hypotheses; 3) no explicit hypothesis label is required; 4) the shared CNN may be well pre-trained with a large-scale single-label image dataset, e.g. ImageNet; and 5) it may naturally output multi-label prediction results. Experimental results on Pascal VOC2007 and VOC2012 multi-label image datasets well demonstrate the superiority of the proposed HCP infrastructure over other state-of-the-arts. In particular, the mAP reaches 84.2% by HCP only and 90.3% after the fusion with our complementary result in [47] based on hand-crafted features on the VOC2012 dataset, which significantly outperforms the state-of-the-arts with a large margin of more than 7%.