Jian Zhu

BigScience Workshop, Teven Le Scao, Angela Fan et al. · allen-ai, berkeley

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

Jian Zhu, Cong Zhang, David Jurgens

In this study, we tackle massively multilingual grapheme-to-phoneme conversion through implementing G2P models based on ByT5. We have curated a G2P dataset from various sources that covers around 100 languages and trained large-scale multilingual G2P models based on ByT5. We found that ByT5 operating on byte-level inputs significantly outperformed the token-based mT5 model in terms of multilingual G2P. Pairwise comparison with monolingual models in these languages suggests that multilingual ByT5 models generally lower the phone error rate by jointly learning from a variety of languages. The pretrained model can further benefit low resource G2P through zero-shot prediction on unseen languages or provides pretrained weights for finetuning, which helps the model converge to a lower phone error rate than randomly initialized weights. To facilitate future research on multilingual G2P, we make available our code and pretrained multilingual G2P models at: https://github.com/lingjzhu/CharsiuG2P.

Hugo Laurençon, Lucile Saulnier, Thomas Wang et al. · huggingface

As language models grow ever larger, the need for large-scale high-quality text datasets has never been more pressing, especially in multilingual settings. The BigScience workshop, a 1-year international and multidisciplinary initiative, was formed with the goal of researching and training large language models as a values-driven undertaking, putting issues of ethics, harm, and governance in the foreground. This paper documents the data creation and curation efforts undertaken by BigScience to assemble the Responsible Open-science Open-collaboration Text Sources (ROOTS) corpus, a 1.6TB dataset spanning 59 languages that was used to train the 176-billion-parameter BigScience Large Open-science Open-access Multilingual (BLOOM) language model. We further release a large initial subset of the corpus and analyses thereof, and hope to empower large-scale monolingual and multilingual modeling projects with both the data and the processing tools, as well as stimulate research around this large multilingual corpus.

Chong Ma, Taiyi Su, Jian Zhu et al.

Vision-language-action (VLA) policies operate in a closed loop in real-world robot tasks: a robot observes the scene, executes an action chunk, and conditions its next decision on the resulting observation. However, most existing world models for robot action evaluation are limited to open-loop prediction along pre-collected action trajectories. This prevents them from supporting closed-loop VLA evaluation, where each action chunk must be conditioned on the observation generated by the previous execution. To address this gap, we propose PiL-World, a chunk-wise world model designed for policy-in-the-loop VLA evaluation. Given the current observation and the action trajectory rolled out by a VLA policy, PiL-World generates multi-view future observations that are consistent with the VLA rollout and match the image inputs required by the policy. By alternating between VLA inference and world-model prediction, PiL-World enables closed-loop evaluation without real robot execution at every step. To improve rollout fidelity, PiL-World conditions video generation on action-derived visual control from head-view robot motion and latent histories that encode task execution context, while jointly predicting complementary multi-view observations. Beyond successful teleoperated demonstrations, it also learns from failed execution trajectories, helping the imagined rollouts better match the distribution of real policy executions. We evaluate PiL-World on three real dual-arm manipulation tasks. PiL-World generates imagined rollouts that are highly consistent with real robot executions. More importantly, compared with the baseline, it reduces the error between VLA success rates measured in real-world rollouts and those estimated through closed-loop world-model evaluation from 63.2% to 12.0%.

Jian Zhu, Zhidong Lin, Wensheng Gan et al.

Discovering valuable insights from rich data is a crucial task for exploratory data analysis. Sequential pattern mining (SPM) has found widespread applications across various domains. In recent years, low-utility sequential pattern mining (LUSPM) has shown strong potential in applications such as intrusion detection and genomic sequence analysis. However, existing research in utility-based SPM focuses on high-utility sequential patterns, and the definitions and strategies used in high-utility SPM cannot be directly applied to LUSPM. Moreover, no algorithms have yet been developed specifically for mining low-utility sequential patterns. To address these problems, we formalize the LUSPM problem, redefine sequence utility, and introduce a compact data structure called the sequence-utility chain to efficiently record utility information. Furthermore, we propose three novel algorithm--LUSPM_b, LUSPM_s, and LUSPM_e--to discover the complete set of low-utility sequential patterns. LUSPM_b serves as an exhaustive baseline, while LUSPM_s and LUSPM_e build upon it, generating subsequences through shrinkage and extension operations, respectively. In addition, we introduce the maximal non-mutually contained sequence set and incorporate multiple pruning strategies, which significantly reduce redundant operations in both LUSPM_s and LUSPM_e. Finally, extensive experimental results demonstrate that both LUSPM_s and LUSPM_e substantially outperform LUSPM_b and exhibit excellent scalability. Notably, LUSPM_e achieves superior efficiency, requiring less runtime and memory consumption than LUSPM_s. Our code is available at https://github.com/Zhidong-Lin/LUSPM.

Shikhar Bharadwaj, Chin-Jou Li, Yoonjae Kim et al.

Phone recognition (PR) serves as the atomic interface for language-agnostic modeling for cross-lingual speech processing and phonetic analysis. Despite prolonged efforts in developing PR systems, current evaluations only measure surface-level transcription accuracy. We introduce PRiSM, the first open-source benchmark designed to expose blind spots in phonetic perception through intrinsic and extrinsic evaluation of PR systems. PRiSM standardizes transcription-based evaluation and assesses downstream utility in clinical, educational, and multilingual settings with transcription and representation probes. We find that diverse language exposure during training is key to PR performance, encoder-CTC models are the most stable, and specialized PR models still outperform Large Audio Language Models. PRiSM releases code, recipes, and datasets to move the field toward multilingual speech models with robust phonetic ability: https://github.com/changelinglab/prism.

Taiyi Su, Jian Zhu, Tianjian Wang et al.

Real-world household robots require Vision-Language-Action (VLA) foundation models that can acquire reusable manipulation skills across diverse objects, task conditions, and household environments. Deformable-object folding is a representative challenge, requiring robots to handle clothing items from random initial states across varying categories, geometries, materials, and scenes. However, existing VLA systems commonly train separate policies for different object categories, while naively mixed multi-task training often suffers from task interference and degraded performance. To move beyond category-specific folding policies, we introduce DeMaVLA, a VLA foundation model for generalizable Deformable Manipulation. DeMaVLA adopts a VLM backbone with an action expert and formulates continuous action generation using flow matching. To improve efficiency, the action expert is constructed by pruning every other transformer layer while preserving layer-wise alignment with the VLM backbone, reducing training and inference cost. DeMaVLA is first pre-trained on approximately 5,000 hours of selected real-world dual-arm demonstrations to acquire general manipulation priors. It is then post-trained on mixed folding data that aggregates self-collected demonstrations and corrective trajectories from real-robot failures across multiple folding tasks through a human-in-the-loop Data Aggregation~(DAgger) pipeline. Experiments show that DeMaVLA achieves competitive performance on RoboTwin and strong real-world results on our household folding benchmark. These results highlight the value of scalable real-world data, efficient action generation, and corrective learning for general-purpose VLA policies in deformable-object manipulation.

Jian Zhu, Hanli Wang, Miaojing Shi

The visual commonsense reasoning (VCR) task is to choose an answer and provide a justifying rationale based on the given image and textural question. Representative works first recognize objects in images and then associate them with key words in texts. However, existing approaches do not consider exact positions of objects in a human-like three-dimensional (3D) manner, making them incompetent to accurately distinguish objects and understand visual relation. Recently, multi-modal large language models (MLLMs) have been used as powerful tools for several multi-modal tasks but not for VCR yet, which requires elaborate reasoning on specific visual objects referred by texts. In light of the above, an MLLM enhanced pseudo 3D perception framework is designed for VCR. Specifically, we first demonstrate that the relation between objects is relevant to object depths in images, and hence introduce object depth into VCR frameworks to infer 3D positions of objects in images. Then, a depth-aware Transformer is proposed to encode depth differences between objects into the attention mechanism of Transformer to discriminatively associate objects with visual scenes guided by depth. To further associate the answer with the depth of visual scene, each word in the answer is tagged with a pseudo depth to realize depth-aware association between answer words and objects. On the other hand, BLIP-2 as an MLLM is employed to process images and texts, and the referring expressions in texts involving specific visual objects are modified with linguistic object labels to serve as comprehensible MLLM inputs. Finally, a parameter optimization technique is devised to fully consider the quality of data batches based on multi-level reasoning confidence. Experiments on the VCR dataset demonstrate the superiority of the proposed framework over state-of-the-art approaches.

Congcong Liu, Yuejiang Li, Jian Zhu et al.

Click-through rate (CTR) Prediction is of great importance in real-world online ads systems. One challenge for the CTR prediction task is to capture the real interest of users from their clicked items, which is inherently biased by presented positions of items, i.e., more front positions tend to obtain higher CTR values. A popular line of existing works focuses on explicitly estimating position bias by result randomization which is expensive and inefficient, or by inverse propensity weighting (IPW) which relies heavily on the quality of the propensity estimation. Another common solution is modeling position as features during offline training and simply adopting fixed value or dropout tricks when serving. However, training-inference inconsistency can lead to sub-optimal performance. Furthermore, post-click information such as position values is informative while less exploited in CTR prediction. This work proposes a simple yet efficient knowledge distillation framework to alleviate the impact of position bias and leverage position information to improve CTR prediction. We demonstrate the performance of our proposed method on a real-world production dataset and online A/B tests, achieving significant improvements over competing baseline models. The proposed method has been deployed in the real world online ads systems, serving main traffic on one of the world's largest e-commercial platforms.

Bo Peng, Daniel Goldstein, Quentin Anthony et al. · harvard

We present Eagle (RWKV-5) and Finch (RWKV-6), sequence models improving upon the RWKV (RWKV-4) architecture. Our architectural design advancements include multi-headed matrix-valued states and a dynamic recurrence mechanism that improve expressivity while maintaining the inference efficiency characteristics of RNNs. We introduce a new multilingual corpus with 1.12 trillion tokens and a fast tokenizer based on greedy matching for enhanced multilinguality. We trained four Eagle models, ranging from 0.46 to 7.5 billion parameters, and two Finch models with 1.6 and 3.1 billion parameters and find that they achieve competitive performance across a wide variety of benchmarks. We release all our models on HuggingFace under the Apache 2.0 license. Models at: https://huggingface.co/RWKV Training code at: https://github.com/RWKV/RWKV-LM Inference code at: https://github.com/RWKV/ChatRWKV Time-parallel training code at: https://github.com/RWKV/RWKV-infctx-trainer

Jian Zhu, Congcong Liu, Pei Wang et al.

Model evolution and constant availability of data are two common phenomena in large-scale real-world machine learning applications, e.g. ads and recommendation systems. To adapt, the real-world system typically retrain with all available data and online learn with recently available data to update the models periodically with the goal of better serving performance. In this paper, we propose a novel framework, named Confidence Ranking, which designs the optimization objective as a ranking function with two different models. Our confidence ranking loss allows direct optimization of the logits output for different convex surrogate functions of metrics, e.g. AUC and Accuracy depending on the target task and dataset. Armed with our proposed methods, our experiments show that the introduction of confidence ranking loss can outperform all baselines on the CTR prediction tasks of public and industrial datasets. This framework has been deployed in the advertisement system of JD.com to serve the main traffic in the fine-rank stage.

Jun Sun, Xinxin Zhang, Simin Hong et al.

Multimodal learning, while contributing to numerous success stories across various fields, faces the challenge of prohibitively expensive manual annotation. To address the scarcity of annotated data, a popular solution is unsupervised domain adaptation, which has been extensively studied in unimodal settings yet remains less explored in multimodal settings. In this paper, we investigate heterogeneous multimodal domain adaptation, where the primary challenge is the varying domain shifts of different modalities from the source to the target domain. We first introduce the information bottleneck method to learn representations for each modality independently, and then match the source and target domains in the representation space with correlation alignment. To balance the domain alignment of all modalities, we formulate the problem as a multi-objective task, aiming for a Pareto optimal solution. By exploiting the properties specific to our model, the problem can be simplified to a quadratic programming problem. Further approximation yields a closed-form solution, leading to an efficient modality-balanced multimodal domain adaptation algorithm. The proposed method features \textbf{B}alanced multi-\textbf{o}bjective \textbf{o}ptimization for \textbf{m}ultimodal \textbf{d}omain \textbf{a}daptation, termed \textbf{Boomda}. Extensive empirical results showcase the effectiveness of the proposed approach and demonstrate that Boomda outperforms the competing schemes. The code is is available at: https://github.com/sunjunaimer/Boomda.git.

Jian Zhu, Xin Zou, Jun Sun et al.

In recent years, the advancement of Graph Neural Networks (GNNs) has significantly propelled progress in Multi-View Clustering (MVC). However, existing methods face the problem of coarse-grained graph fusion. Specifically, current approaches typically generate a separate graph structure for each view and then perform weighted fusion of graph structures at the view level, which is a relatively rough strategy. To address this limitation, we present a novel Mixture of Ego-Graphs Contrastive Representation Learning (MoEGCL). It mainly consists of two modules. In particular, we propose an innovative Mixture of Ego-Graphs Fusion (MoEGF), which constructs ego graphs and utilizes a Mixture-of-Experts network to implement fine-grained fusion of ego graphs at the sample level, rather than the conventional view-level fusion. Additionally, we present the Ego Graph Contrastive Learning (EGCL) module to align the fused representation with the view-specific representation. The EGCL module enhances the representation similarity of samples from the same cluster, not merely from the same sample, further boosting fine-grained graph representation. Extensive experiments demonstrate that MoEGCL achieves state-of-the-art results in deep multi-view clustering tasks. The source code is publicly available at https://github.com/HackerHyper/MoEGCL.

Jian Zhu, Zuoyu Tian, Yadong Liu et al.

Inducing semantic representations directly from speech signals is a highly challenging task but has many useful applications in speech mining and spoken language understanding. This study tackles the unsupervised learning of semantic representations for spoken utterances. Through converting speech signals into hidden units generated from acoustic unit discovery, we propose WavEmbed, a multimodal sequential autoencoder that predicts hidden units from a dense representation of speech. Secondly, we also propose S-HuBERT to induce meaning through knowledge distillation, in which a sentence embedding model is first trained on hidden units and passes its knowledge to a speech encoder through contrastive learning. The best performing model achieves a moderate correlation (0.5~0.6) with human judgments, without relying on any labels or transcriptions. Furthermore, these models can also be easily extended to leverage textual transcriptions of speech to learn much better speech embeddings that are strongly correlated with human annotations. Our proposed methods are applicable to the development of purely data-driven systems for speech mining, indexing and search.

Jian Zhu, Zhangmin Huang, Xiaohu Ruan et al.

Learning the hash representation of multi-view heterogeneous data is an important task in multimedia retrieval. However, existing methods fail to effectively fuse the multi-view features and utilize the metric information provided by the dissimilar samples, leading to limited retrieval precision. Current methods utilize weighted sum or concatenation to fuse the multi-view features. We argue that these fusion methods cannot capture the interaction among different views. Furthermore, these methods ignored the information provided by the dissimilar samples. We propose a novel deep metric multi-view hashing (DMMVH) method to address the mentioned problems. Extensive empirical evidence is presented to show that gate-based fusion is better than typical methods. We introduce deep metric learning to the multi-view hashing problems, which can utilize metric information of dissimilar samples. On the MIR-Flickr25K, MS COCO, and NUS-WIDE, our method outperforms the current state-of-the-art methods by a large margin (up to 15.28 mean Average Precision (mAP) improvement).

Jian Zhu, Yu Cui, Zhangmin Huang et al.

The multi-view hash method converts heterogeneous data from multiple views into binary hash codes, which is one of the critical technologies in multimedia retrieval. However, the current methods mainly explore the complementarity among multiple views while lacking confidence learning and fusion. Moreover, in practical application scenarios, the single-view data contain redundant noise. To conduct the confidence learning and eliminate unnecessary noise, we propose a novel Adaptive Confidence Multi-View Hashing (ACMVH) method. First, a confidence network is developed to extract useful information from various single-view features and remove noise information. Furthermore, an adaptive confidence multi-view network is employed to measure the confidence of each view and then fuse multi-view features through a weighted summation. Lastly, a dilation network is designed to further enhance the feature representation of the fused features. To the best of our knowledge, we pioneer the application of confidence learning into the field of multimedia retrieval. Extensive experiments on two public datasets show that the proposed ACMVH performs better than state-of-the-art methods (maximum increase of 3.24%). The source code is available at https://github.com/HackerHyper/ACMVH.

Jian Zhu, Wen Cheng, Yu Cui et al.

Hash representation learning of multi-view heterogeneous data is the key to improving the accuracy of multimedia retrieval. However, existing methods utilize local similarity and fall short of deeply fusing the multi-view features, resulting in poor retrieval accuracy. Current methods only use local similarity to train their model. These methods ignore global similarity. Furthermore, most recent works fuse the multi-view features via a weighted sum or concatenation. We contend that these fusion methods are insufficient for capturing the interaction between various views. We present a novel Central Similarity Multi-View Hashing (CSMVH) method to address the mentioned problems. Central similarity learning is used for solving the local similarity problem, which can utilize the global similarity between the hash center and samples. We present copious empirical data demonstrating the superiority of gate-based fusion over conventional approaches. On the MS COCO and NUS-WIDE, the proposed CSMVH performs better than the state-of-the-art methods by a large margin (up to 11.41% mean Average Precision (mAP) improvement).

Jian Zhu, Changbing Yang, Farhan Samir et al.

In this project, we demonstrate that phoneme-based models for speech processing can achieve strong crosslinguistic generalizability to unseen languages. We curated the IPAPACK, a massively multilingual speech corpora with phonemic transcriptions, encompassing more than 115 languages from diverse language families, selectively checked by linguists. Based on the IPAPACK, we propose CLAP-IPA, a multi-lingual phoneme-speech contrastive embedding model capable of open-vocabulary matching between arbitrary speech signals and phonemic sequences. The proposed model was tested on 95 unseen languages, showing strong generalizability across languages. Temporal alignments between phonemes and speech signals also emerged from contrastive training, enabling zeroshot forced alignment in unseen languages. We further introduced a neural forced aligner IPA-ALIGNER by finetuning CLAP-IPA with the Forward-Sum loss to learn better phone-to-audio alignment. Evaluation results suggest that IPA-ALIGNER can generalize to unseen languages without adaptation.

Anton Lozhkov, Raymond Li, Loubna Ben Allal et al. · berkeley, ibm-research

The BigCode project, an open-scientific collaboration focused on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder2. In partnership with Software Heritage (SWH), we build The Stack v2 on top of the digital commons of their source code archive. Alongside the SWH repositories spanning 619 programming languages, we carefully select other high-quality data sources, such as GitHub pull requests, Kaggle notebooks, and code documentation. This results in a training set that is 4x larger than the first StarCoder dataset. We train StarCoder2 models with 3B, 7B, and 15B parameters on 3.3 to 4.3 trillion tokens and thoroughly evaluate them on a comprehensive set of Code LLM benchmarks. We find that our small model, StarCoder2-3B, outperforms other Code LLMs of similar size on most benchmarks, and also outperforms StarCoderBase-15B. Our large model, StarCoder2- 15B, significantly outperforms other models of comparable size. In addition, it matches or outperforms CodeLlama-34B, a model more than twice its size. Although DeepSeekCoder- 33B is the best-performing model at code completion for high-resource languages, we find that StarCoder2-15B outperforms it on math and code reasoning benchmarks, as well as several low-resource languages. We make the model weights available under an OpenRAIL license and ensure full transparency regarding the training data by releasing the SoftWare Heritage persistent IDentifiers (SWHIDs) of the source code data.

Wenjing Hong, Zhonghua Rong, Li Wang et al.

Large Language Models (LLMs) have been widely deployed, especially through free Web-based applications that expose them to diverse user-generated inputs, including those from long-tail distributions such as low-resource languages and encrypted private data. This open-ended exposure increases the risk of jailbreak attacks that undermine model safety alignment. While recent studies have shown that leveraging long-tail distributions can facilitate such jailbreaks, existing approaches largely rely on handcrafted rules, limiting the systematic evaluation of these security and privacy vulnerabilities. In this work, we present EvoJail, an automated framework for discovering long-tail distribution attacks via multi-objective evolutionary search. EvoJail formulates long-tail attack prompt generation as a multi-objective optimization problem that jointly maximizes attack effectiveness and minimizes output perplexity, and introduces a semantic-algorithmic solution representation to capture both high-level semantic intent and low-level structural transformations of encryption-decryption logic. Building upon this representation, EvoJail integrates LLM-assisted operators into a multi-objective evolutionary framework, enabling adaptive and semantically informed mutation and crossover for efficiently exploring a highly structured and open-ended search space. Extensive experiments demonstrate that EvoJail consistently discovers diverse and effective long-tail jailbreak strategies, achieving competitive performance with existing methods in both individual and ensemble level.

Jian Zhu, Ilya Valmianski, Anitha Kannan

AI-driven medical history-taking is an important component in symptom checking, automated patient intake, triage, and other AI virtual care applications. As history-taking is extremely varied, machine learning models require a significant amount of data to train. To overcome this challenge, existing systems are developed using indirect data or expert knowledge. This leads to a training-inference gap as models are trained on different kinds of data than what they observe at inference time. In this work, we present a two-stage re-ranking approach that helps close the training-inference gap by re-ranking the first-stage question candidates using a dialogue-contextualized model. For this, we propose a new model, global re-ranker, which cross-encodes the dialogue with all questions simultaneously, and compare it with several existing neural baselines. We test both transformer and S4-based language model backbones. We find that relative to the expert system, the best performance is achieved by our proposed global re-ranker with a transformer backbone, resulting in a 30% higher normalized discount cumulative gain (nDCG) and a 77% higher mean average precision (mAP).

Xiao Yu Cindy Zhang, Carlos R. Ferreira, Francis Rossignol et al.

Rare diseases, including Inborn Errors of Metabolism (IEM), pose significant diagnostic challenges. Case reports serve as key but computationally underutilized resources to inform diagnosis. Clinical dense information extraction refers to organizing medical information into structured predefined categories. Large Language Models (LLMs) may enable scalable information extraction from case reports but are rarely evaluated for this task. We introduce CaseReportBench, an expert-annotated dataset for dense information extraction of case reports, focusing on IEMs. Using this dataset, we assess various models and prompting strategies, introducing novel approaches such as category-specific prompting and subheading-filtered data integration. Zero-shot chain-of-thought prompting offers little advantage over standard zero-shot prompting. Category-specific prompting improves alignment with the benchmark. The open-source model Qwen2.5-7B outperforms GPT-4o for this task. Our clinician evaluations show that LLMs can extract clinically relevant details from case reports, supporting rare disease diagnosis and management. We also highlight areas for improvement, such as LLMs' limitations in recognizing negative findings important for differential diagnosis. This work advances LLM-driven clinical natural language processing and paves the way for scalable medical AI applications.

Jian Zhu, He Wang, Yang Xu et al.

Hyperspectral and multispectral image (HSI-MSI) fusion involves combining a low-resolution hyperspectral image (LR-HSI) with a high-resolution multispectral image (HR-MSI) to generate a high-resolution hyperspectral image (HR-HSI). Most deep learning-based methods for HSI-MSI fusion rely on large amounts of hyperspectral data for supervised training, which is often scarce in practical applications. In this paper, we propose a self-learning Adaptive Residual Guided Subspace Diffusion Model (ARGS-Diff), which only utilizes the observed images without any extra training data. Specifically, as the LR-HSI contains spectral information and the HR-MSI contains spatial information, we design two lightweight spectral and spatial diffusion models to separately learn the spectral and spatial distributions from them. Then, we use these two models to reconstruct HR-HSI from two low-dimensional components, i.e, the spectral basis and the reduced coefficient, during the reverse diffusion process. Furthermore, we introduce an Adaptive Residual Guided Module (ARGM), which refines the two components through a residual guided function at each sampling step, thereby stabilizing the sampling process. Extensive experimental results demonstrate that ARGS-Diff outperforms existing state-of-the-art methods in terms of both performance and computational efficiency in the field of HSI-MSI fusion. Code is available at https://github.com/Zhu1116/ARGS-Diff.

Jian Zhu, Xin Zou, Lei Liu et al.

Multi-view clustering can partition data samples into their categories by learning a consensus representation in an unsupervised way and has received more and more attention in recent years. However, there is an untrusted fusion problem. The reasons for this problem are as follows: 1) The current methods ignore the presence of noise or redundant information in the view; 2) The similarity of contrastive learning comes from the same sample rather than the same cluster in deep multi-view clustering. It causes multi-view fusion in the wrong direction. This paper proposes a novel multi-view clustering network to address this problem, termed as Trusted Mamba Contrastive Network (TMCN). Specifically, we present a new Trusted Mamba Fusion Network (TMFN), which achieves a trusted fusion of multi-view data through a selective mechanism. Moreover, we align the fused representation and the view-specific representation using the Average-similarity Contrastive Learning (AsCL) module. AsCL increases the similarity of view presentation from the same cluster, not merely from the same sample. Extensive experiments show that the proposed method achieves state-of-the-art results in deep multi-view clustering tasks. The source code is available at https://github.com/HackerHyper/TMCN.

Jian Zhu, Mingkai Sheng, Mingda Ke et al.

The multi-modal hashing method is widely used in multimedia retrieval. It can fuse multi-source data to generate binary hash code. However, the current multi-modal methods have the problem of low retrieval accuracy. The reason is that the individual backbone networks have limited feature expression capabilities and are not jointly pre-trained on large-scale unsupervised multi-modal data. To solve this problem, we propose a new baseline CLIP Multi-modal Hashing (CLIPMH) method. It uses CLIP model to extract text and image features, and then fuse to generate hash code. CLIP improves the expressiveness of each modal feature. In this way, it can greatly improve the retrieval performance of multi-modal hashing methods. In comparison to state-of-the-art unsupervised and supervised multi-modal hashing methods, experiments reveal that the proposed CLIPMH can significantly enhance performance (Maximum increase of 8.38%). CLIP also has great advantages over the text and visual backbone networks commonly used before.

Yuchen Luo, Fangyue Zhu, Ruining Zhou et al.

Post-Training Quantization (PTQ) is crucial for efficient model deployment, yet its effectiveness on Ascend NPU remains under-explored compared to GPU architectures. This paper presents a case study of representative PTQ baselines applied to reasoning-oriented models such as DeepSeek-R1-Distill-Qwen series (1.5B/7B/14B) and QwQ-32B. We evaluate four distinct algorithms, including AWQ, GPTQ, SmoothQuant, and FlatQuant, to cover the spectrum from weight-only compression to advanced rotation-based methods. Our empirical results reveal significant platform sensitivity. While 4-bit weight-only quantization proves viable for larger models, aggressive 4-bit weight-activation schemes suffer from layer-wise calibration instability on the NPU, leading to logic collapse in long-context reasoning tasks. Conversely, standard 8-bit quantization remains numerically stable. Furthermore, a real-world INT8 deployment demonstrates that although optimized kernels reduce latency, dynamic quantization overheads currently limit end-to-end acceleration. These findings offer a practical reference for the feasibility and limitations of deploying quantized reasoning models on Ascend NPU.

Jian Zhu, Xin Zou, Xi Wang et al.

In recent years, Multi-View Clustering (MVC) has been significantly advanced under the influence of deep learning. By integrating heterogeneous data from multiple views, MVC enhances clustering analysis, making multi-view fusion critical to clustering performance. However, there is a problem of low-quality data in multi-view fusion. This problem primarily arises from two reasons: 1) Certain views are contaminated by noisy data. 2) Some views suffer from missing data. This paper proposes a novel Stochastic Generative Diffusion Fusion (SGDF) method to address this problem. SGDF leverages a multiple generative mechanism for the multi-view feature of each sample. It is robust to low-quality data. Building on SGDF, we further present the Generative Diffusion Contrastive Network (GDCN). Extensive experiments show that GDCN achieves the state-of-the-art results in deep MVC tasks. The source code is publicly available at https://github.com/HackerHyper/GDCN.

Kenan Alkiek, Anna Wegmann, Jian Zhu et al.

Linguistic style is pivotal for understanding how texts convey meaning and fulfill communicative purposes, yet extracting detailed stylistic features at scale remains challenging. We present Neurobiber, a transformer-based system for fast, interpretable style profiling built on Biber's Multidimensional Analysis (MDA). Neurobiber predicts 96 Biber-style features from our open-source BiberPlus library (a Python toolkit that computes stylistic features and provides integrated analytics, e.g., PCA and factor analysis). Despite being up to 56 times faster than existing open source systems, Neurobiber replicates classic MDA insights on the CORE corpus and achieves competitive performance on the PAN 2020 authorship verification task without extensive retraining. Its efficient and interpretable representations readily integrate into downstream NLP pipelines, facilitating large-scale stylometric research, forensic analysis, and real-time text monitoring. All components are made publicly available.

Yongwei Nie, Mingxian Fan, Chengjiang Long et al.

Human Mesh Recovery (HMR) is the task of estimating a parameterized 3D human mesh from an image. There is a kind of methods first training a regression model for this problem, then further optimizing the pretrained regression model for any specific sample individually at test time. However, the pretrained model may not provide an ideal optimization starting point for the test-time optimization. Inspired by meta-learning, we incorporate the test-time optimization into training, performing a step of test-time optimization for each sample in the training batch before really conducting the training optimization over all the training samples. In this way, we obtain a meta-model, the meta-parameter of which is friendly to the test-time optimization. At test time, after several test-time optimization steps starting from the meta-parameter, we obtain much higher HMR accuracy than the test-time optimization starting from the simply pretrained regression model. Furthermore, we find test-time HMR objectives are different from training-time objectives, which reduces the effectiveness of the learning of the meta-model. To solve this problem, we propose a dual-network architecture that unifies the training-time and test-time objectives. Our method, armed with meta-learning and the dual networks, outperforms state-of-the-art regression-based and optimization-based HMR approaches, as validated by the extensive experiments. The codes are available at https://github.com/fmx789/Meta-HMR.

Jian Zhu, Cong Zhang, David Jurgens

The task of phone-to-audio alignment has many applications in speech research. Here we introduce two Wav2Vec2-based models for both text-dependent and text-independent phone-to-audio alignment. The proposed Wav2Vec2-FS, a semi-supervised model, directly learns phone-to-audio alignment through contrastive learning and a forward sum loss, and can be coupled with a pretrained phone recognizer to achieve text-independent alignment. The other model, Wav2Vec2-FC, is a frame classification model trained on forced aligned labels that can both perform forced alignment and text-independent segmentation. Evaluation results suggest that both proposed methods, even when transcriptions are not available, generate highly close results to existing forced alignment tools. Our work presents a neural pipeline of fully automated phone-to-audio alignment. Code and pretrained models are available at https://github.com/lingjzhu/charsiu.

Jian Zhu, Will Styler, Ian Calloway

For speech research, ultrasound tongue imaging provides a non-invasive means for visualizing tongue position and movement during articulation. Extracting tongue contours from ultrasound images is a basic step in analyzing ultrasound data but this task often requires non-trivial manual annotation. This study presents an open source tool for fully automatic tracking of tongue contours in ultrasound frames using neural network based methods. We have implemented and systematically compared two convolutional neural networks, U-Net and DenseU-Net, under different conditions. Though both models can perform automatic contour tracking with comparable accuracy, Dense U-Net architecture seems more generalizable across test datasets while U-Net has faster extraction speed. Our comparison also shows that the choice of loss function and data augmentation have a greater effect on tracking performance in this task. This public available segmentation tool shows considerable promise for the automated tongue contour annotation of ultrasound images in speech research.

Chaojie Zhao, Peng Zhang, Jian Zhu et al.

A challenge in speech production research is to predict future tongue movements based on a short period of past tongue movements. This study tackles speaker-dependent tongue motion prediction problem in unlabeled ultrasound videos with convolutional long short-term memory (ConvLSTM) networks. The model has been tested on two different ultrasound corpora. ConvLSTM outperforms 3-dimensional convolutional neural network (3DCNN) in predicting the 9\textsuperscript{th} frames based on 8 preceding frames, and also demonstrates good capacity to predict only the tongue contours in future frames. Further tests reveal that ConvLSTM can also learn to predict tongue movements in more distant frames beyond the immediately following frames. Our codes are available at: https://github.com/shuiliwanwu/ConvLstm-ultrasound-videos.

Jingmou Xian, Jian Zhu, Haolin Liao et al.

Sparse-view tomographic reconstruction is a pivotal direction for reducing radiation dose and augmenting clinical applicability. While many research works have proposed the reconstruction of tomographic images from sparse 2D projections, existing models tend to excessively focus on high-frequency information while overlooking low-frequency components within the sparse input images. This bias towards high-frequency information often leads to overfitting, particularly intense at edges and boundaries in the reconstructed slices. In this paper, we introduce the Frequency Regularized Neural Attenuation/Activity Field (Freq-NAF) for self-supervised sparse-view tomographic reconstruction. Freq-NAF mitigates overfitting by incorporating frequency regularization, directly controlling the visible frequency bands in the neural network input. This approach effectively balances high-frequency and low-frequency information. We conducted numerical experiments on CBCT and SPECT datasets, and our method demonstrates state-of-the-art accuracy.

Robin Shing-Hei Yuen, Timothy Tin-Long Tse, Jian Zhu

Current speech-based LLMs are predominantly trained on extensive ASR and TTS datasets, excelling in tasks related to these domains. However, their ability to handle direct speech-to-speech conversations remains notably constrained. These models often rely on an ASR-to-TTS chain-of-thought pipeline, converting speech into text for processing before generating audio responses, which introduces latency and loses audio features. We propose a method that implicitly internalizes ASR chain of thought into a speech LLM, enhancing its native speech understanding capabilities. Our approach reduces latency and improves the model's native understanding of speech, paving the way for more efficient and natural real-time audio interactions. We also release a large-scale synthetic conversational dataset to facilitate further research.

Changbing Yang, Franklin Ma, Freda Shi et al.

This paper introduces LingGym, a new benchmark that evaluates LLMs' capacity for meta-linguistic reasoning using Interlinear Glossed Text (IGT) and grammatical descriptions extracted from 18 typologically diverse reference grammars. Unlike previous work that focuses on specific downstream tasks, we assess whether LLMs can generalize linguistic inference across low-resource languages and structures not seen during training. We present a controlled evaluation task: Word-Gloss Inference, in which the model must infer a missing word and gloss from context using varying levels of linguistic information (e.g., glosses, grammatical explanations, translations). Our results show that incorporating structured linguistic cues leads to consistent improvements in reasoning performance across all models. This work highlights both the promise and current limitations of using LLMs for typologically informed linguistic analysis and low-resource language documentation.

Ganxi Xu, Zhao-Rong Lai, Yuting Tang et al.

Brain encoding models not only serve to decipher how visual stimuli are transformed into neural responses, but also represent a critical step toward visual prostheses that restore vision for patients with severe vision disorders. Brain encoding involves two fundamental steps: achieving faithful reconstruction of neural responses and establishing cross-modal alignment between visual stimuli and neural responses. To this end, we propose ViBE, a novel brain encoding framework for generating magnetoencephalography (MEG) and electroencephalography (EEG) signals from visual stimuli. Specifically, we first design a spatio-temporal convolutional variational autoencoder (TSC-VAE) that captures the spatio-temporal characteristics of M/EEG signals for effective neural response reconstruction. To bridge the modality gap between visual features and neural representations, we employ Q-Former to map CLIP image embeddings to the TSC-VAE latent space, producing neural proxy embeddings. For comprehensive cross-modal alignment, we combine mean squared error (MSE) loss for point-wise feature matching with sliced Wasserstein distance (SWD) for probability distribution alignment between the neural proxy embeddings and TSC-VAE latent embeddings. We conduct extensive experiments on the THINGS-EEG2 and THINGS-MEG datasets, demonstrating the effectiveness of our approach in generating high-quality M/EEG signals from visual stimuli.

Rongze Liu, Jiaxin Pei, Jian Zhu

Anime, originated from Japan, is one of the most influential cultural products in modern society and is especially popular among younger generations. The popularity of anime reflects important cultural evolutions in our society. Despite existing research on anime as a cultural phenomenon, we still have a limited understanding of how anime really evolves over the years. In this study, using a large-scale multimodal dataset of anime characters from an anime review site, we applied computational methods that integrate textual, visual, and production features of anime characters with online popularity traces. By combining LLM-extracted personality features with avatar features, we identify recurring personality archetypes and visual tropes with their temporal evolution over the past decades. We found that the target audience of anime has undergone a systematic shift from children to a maturing audience of teenagers and young adults over time. Character design has been undergoing moe-ification, with softer or sexualized female traits becoming increasingly prominent since the 2000s. Some personality archetypes are often visually predictable, yet audiences also tend to prefer less conventionalized characters. Finally, we reveal that visual signals play a more dominant role than personality traits in shaping audience preferences, with features such as moe-style faces and mechanical designs contributing greatly to popularity. These findings offer insights into the broader dynamics of anime's cultural and creative practices.

Xiang Chen, Hao Li, Jiangxin Dong et al.

This paper presents a review for the LoViF Challenge on Real-World All-in-One Image Restoration. The challenge aimed to advance research on real-world all-in-one image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow. It provided a unified benchmark to evaluate the robustness and generalization ability of restoration models across multiple degradation categories within a common framework. The competition attracted 124 registered participants and received 9 valid final submissions with corresponding fact sheets, significantly contributing to the progress of real-world all-in-one image restoration. This report provides a detailed analysis of the submitted methods and corresponding results, emphasizing recent progress in unified real-world image restoration. The analysis highlights effective approaches and establishes a benchmark for future research in real-world low-level vision.

Jian Zhu, Zhengyu Jia, Tian Gao et al.

Advanced end-to-end autonomous driving systems predict other vehicles' motions and plan ego vehicle's trajectory. The world model that can foresee the outcome of the trajectory has been used to evaluate the autonomous driving system. However, existing world models predominantly emphasize the trajectory of the ego vehicle and leave other vehicles uncontrollable. This limitation hinders their ability to realistically simulate the interaction between the ego vehicle and the driving scenario. In this paper, we propose a driving World Model named EOT-WM, unifying Ego-Other vehicle Trajectories in videos for driving simulation. Specifically, it remains a challenge to match multiple trajectories in the BEV space with each vehicle in the video to control the video generation. We first project ego-other vehicle trajectories in the BEV space into the image coordinate for vehicle-trajectory match via pixel positions. Then, trajectory videos are encoded by the Spatial-Temporal Variational Auto Encoder to align with driving video latents spatially and temporally in the unified visual space. A trajectory-injected diffusion Transformer is further designed to denoise the noisy video latents for video generation with the guidance of ego-other vehicle trajectories. In addition, we propose a metric based on control latent similarity to evaluate the controllability of trajectories. Extensive experiments are conducted on the nuScenes dataset, and the proposed model outperforms the state-of-the-art method by 30% in FID and 55% in FVD. The model can also predict unseen driving scenes with self-produced trajectories.

Jian Zhu, Farhan Samir, Eleanor Chodroff et al. · cmu

We present ZIPA, a family of efficient speech models that advances the state-of-the-art performance of crosslinguistic phone recognition. We first curated IPAPack++, a large-scale multilingual speech corpus with 17,132 hours of normalized phone transcriptions and a novel evaluation set capturing unseen languages and sociophonetic variation. With the large-scale training data, ZIPA, including transducer (ZIPA-T) and CTC-based (ZIPA-CR) variants, leverage the efficient Zipformer backbones and outperform existing phone recognition systems with much fewer parameters. Further scaling via noisy student training on 11,000 hours of pseudo-labeled multilingual data yields further improvement. While ZIPA achieves strong performance on benchmarks, error analysis reveals persistent limitations in modeling sociophonetic diversity, underscoring challenges for future research.

Shenran Wang, Changbing Yang, Mike Parkhill et al.

We present lightweight flow matching multilingual text-to-speech (TTS) systems for Ojibwe, Mi'kmaq, and Maliseet, three Indigenous languages in North America. Our results show that training a multilingual TTS model on three typologically similar languages can improve the performance over monolingual models, especially when data are scarce. Attention-free architectures are highly competitive with self-attention architecture with higher memory efficiency. Our research not only advances technical development for the revitalization of low-resource languages but also highlights the cultural gap in human evaluation protocols, calling for a more community-centered approach to human evaluation.

Jian Zhu, Yuping Ruan, Jingfei Chang et al.

The detection of abusive language remains a long-standing challenge with the extensive use of social networks. The detection task of abusive language suffers from limited accuracy. We argue that the existing detection methods utilize the fine-tuning technique of the pre-trained language models (PLMs) to handle downstream tasks. Hence, these methods fail to stimulate the general knowledge of the PLMs. To address the problem, we propose a novel Deep Prompt Multi-task Network (DPMN) for abuse language detection. Specifically, DPMN first attempts to design two forms of deep prompt tuning and light prompt tuning for the PLMs. The effects of different prompt lengths, tuning strategies, and prompt initialization methods on detecting abusive language are studied. In addition, we propose a Task Head based on Bi-LSTM and FFN, which can be used as a short text classifier. Eventually, DPMN utilizes multi-task learning to improve detection metrics further. The multi-task network has the function of transferring effective knowledge. The proposed DPMN is evaluated against eight typical methods on three public datasets: OLID, SOLID, and AbuseAnalyzer. The experimental results show that our DPMN outperforms the state-of-the-art methods.

Jian Zhu, Jianwei Cui, Shihao Chen et al.

We present AccompGen, a system that generates instrumental music audio to accompany input vocals. Given isolated singing voice, AccompGen produces a coherent instrumental accompaniment that can be directly mixed with the input to create complete music. We propose three key innovations over prior work: (1) a dual-rate codec tokenization scheme using HuBERT semantic tokens at 50,Hz for vocals and EnCodec acoustic tokens at 75,Hz for instrumentals, enabling time-aligned yet rate-independent modeling; (2) a three-stage hierarchical autoregressive architecture (semantic to coarse acoustic to fine acoustic) with interleaved multi-codebook prediction and classifier-free guidance; and (3) modern Transformer design choices including QK-norm, GEGLU activations, RMSNorm, and T5-style relative position bias for improved training stability and sequence generalization.

Taiyi Su, Jian Zhu, Yaxuan Li et al.

Embodied world models aim to predict and interact with the physical world through visual observations and actions. However, existing models struggle to accurately translate low-level actions (e.g., joint positions) into precise robotic movements in predicted frames, leading to inconsistencies with real-world physical interactions. To address these limitations, we propose MTV-World, an embodied world model that introduces Multi-view Trajectory-Video control for precise visuomotor prediction. Specifically, instead of directly using low-level actions for control, we employ trajectory videos obtained through camera intrinsic and extrinsic parameters and Cartesian-space transformation as control signals. However, projecting 3D raw actions onto 2D images inevitably causes a loss of spatial information, making a single view insufficient for accurate interaction modeling. To overcome this, we introduce a multi-view framework that compensates for spatial information loss and ensures high-consistency with physical world. MTV-World forecasts future frames based on multi-view trajectory videos as input and conditioning on an initial frame per view. Furthermore, to systematically evaluate both robotic motion precision and object interaction accuracy, we develop an auto-evaluation pipeline leveraging multimodal large models and referring video object segmentation models. To measure spatial consistency, we formulate it as an object location matching problem and adopt the Jaccard Index as the evaluation metric. Extensive experiments demonstrate that MTV-World achieves precise control execution and accurate physical interaction modeling in complex dual-arm scenarios.

Chin-Jou Li, Kalvin Chang, Shikhar Bharadwaj et al. · cmu

Recent advances in spoken language processing have led to substantial progress in phonetic tasks such as automatic speech recognition (ASR), phone recognition (PR), grapheme-to-phoneme conversion (G2P), and phoneme-to-grapheme conversion (P2G). Despite their conceptual similarity, these tasks have largely been studied in isolation, each relying on task-specific architectures and datasets. In this paper, we introduce POWSM (Phonetic Open Whisper-style Speech Model), the first unified framework capable of jointly performing multiple phone-related tasks. POWSM enables seamless conversion between audio, text (graphemes), and phones, opening up new possibilities for universal and low-resource speech processing. Our model outperforms or matches specialized PR models of similar size (Wav2Vec2Phoneme and ZIPA) while jointly supporting G2P, P2G, and ASR. Our training data, code and models are released to foster open science.

Shenran Wang, Timothy Tin-Long Tse, Jian Zhu

We perform in-depth evaluations of in-context learning (ICL) on state-of-the-art transformer, state-space, and hybrid large language models over two categories of knowledge-based ICL tasks. Using a combination of behavioral probing and intervention-based methods, we have discovered that, while LLMs of different architectures can behave similarly in task performance, their internals could remain different. We discover that function vectors (FVs) responsible for ICL are primarily located in the self-attention and Mamba layers, and speculate that Mamba2 uses a different mechanism from FVs to perform ICL. FVs are more important for ICL involving parametric knowledge retrieval, but not for contextual knowledge understanding. Our work contributes to a more nuanced understanding across architectures and task types. Methodologically, our approach also highlights the importance of combining both behavioural and mechanistic analyses to investigate LLM capabilities.

Yunhao Wang, Ziting Li, Shuai Chen et al.

Aligning large-scale vision-language models (VLMs) for complex reasoning via reinforcement learning is often hampered by the limitations of existing policy optimization algorithms, such as static training schedules and the rigid, uniform clipping mechanism in Proximal Policy Optimization (PPO). In this work, we introduce Adaptive Curriculum Policy Optimization (ACPO), a novel framework that addresses these challenges through a dual-component adaptive learning strategy. First, ACPO employs a dynamic curriculum that orchestrates a principled transition from a stable, near on-policy exploration phase to an efficient, off-policy exploitation phase by progressively increasing sample reuse. Second, we propose an Advantage-Aware Adaptive Clipping (AAAC) mechanism that replaces the fixed clipping hyperparameter with dynamic, sample-wise bounds modulated by the normalized advantage of each token. This allows for more granular and robust policy updates, enabling larger gradients for high-potential samples while safeguarding against destructive ones. We conduct extensive experiments on a suite of challenging multimodal reasoning benchmarks, including MathVista, LogicVista, and MMMU-Pro. Results demonstrate that ACPO consistently outperforms strong baselines such as DAPO and PAPO, achieving state-of-the-art performance, accelerated convergence, and superior training stability.

Yue Gong, Shanyuan Liu, Liuzhuozheng Li et al.

We proposed the Chinese Text Adapter-Flux (CTA-Flux). An adaptation method fits the Chinese text inputs to Flux, a powerful text-to-image (TTI) generative model initially trained on the English corpus. Despite the notable image generation ability conditioned on English text inputs, Flux performs poorly when processing non-English prompts, particularly due to linguistic and cultural biases inherent in predominantly English-centric training datasets. Existing approaches, such as translating non-English prompts into English or finetuning models for bilingual mappings, inadequately address culturally specific semantics, compromising image authenticity and quality. To address this issue, we introduce a novel method to bridge Chinese semantic understanding with compatibility in English-centric TTI model communities. Existing approaches relying on ControlNet-like architectures typically require a massive parameter scale and lack direct control over Chinese semantics. In comparison, CTA-flux leverages MultiModal Diffusion Transformer (MMDiT) to control the Flux backbone directly, significantly reducing the number of parameters while enhancing the model's understanding of Chinese semantics. This integration significantly improves the generation quality and cultural authenticity without extensive retraining of the entire model, thus maintaining compatibility with existing text-to-image plugins such as LoRA, IP-Adapter, and ControlNet. Empirical evaluations demonstrate that CTA-flux supports Chinese and English prompts and achieves superior image generation quality, visual realism, and faithful depiction of Chinese semantics.

Shanyuan Liu, Jian Zhu, Junda Lu et al.

Diffusion Transformers (DiTs) have demonstrated exceptional capabilities in text-to-image synthesis. However, in the domain of controllable text-to-image generation using DiTs, most existing methods still rely on the ControlNet paradigm originally designed for UNet-based diffusion models. This paradigm introduces significant parameter overhead and increased computational costs. To address these challenges, we propose the Nano Control Diffusion Transformer (NanoControl), which employs Flux as the backbone network. Our model achieves state-of-the-art controllable text-to-image generation performance while incurring only a 0.024\% increase in parameter count and a 0.029\% increase in GFLOPs, thus enabling highly efficient controllable generation. Specifically, rather than duplicating the DiT backbone for control, we design a LoRA-style (low-rank adaptation) control module that directly learns control signals from raw conditioning inputs. Furthermore, we introduce a KV-Context Augmentation mechanism that integrates condition-specific key-value information into the backbone in a simple yet highly effective manner, facilitating deep fusion of conditional features. Extensive benchmark experiments demonstrate that NanoControl significantly reduces computational overhead compared to conventional control approaches, while maintaining superior generation quality and achieving improved controllability.

Jian Zhu, Shanyuan Liu, Liuzhuozheng Li et al.

Makeup transfer aims to apply the makeup style from a reference face to a target face and has been increasingly adopted in practical applications. Existing GAN-based approaches typically rely on carefully designed loss functions to balance transfer quality and facial identity consistency, while diffusion-based methods often depend on additional face-control modules or algorithms to preserve identity. However, these auxiliary components tend to introduce extra errors, leading to suboptimal transfer results. To overcome these limitations, we propose FLUX-Makeup, a high-fidelity, identity-consistent, and robust makeup transfer framework that eliminates the need for any auxiliary face-control components. Instead, our method directly leverages source-reference image pairs to achieve superior transfer performance. Specifically, we build our framework upon FLUX-Kontext, using the source image as its native conditional input. Furthermore, we introduce RefLoRAInjector, a lightweight makeup feature injector that decouples the reference pathway from the backbone, enabling efficient and comprehensive extraction of makeup-related information. In parallel, we design a robust and scalable data generation pipeline to provide more accurate supervision during training. The paired makeup datasets produced by this pipeline significantly surpass the quality of all existing datasets. Extensive experiments demonstrate that FLUX-Makeup achieves state-of-the-art performance, exhibiting strong robustness across diverse scenarios.