Recent research has investigated the shape and texture biases of deep neural networks (DNNs) in image classification which influence their generalization capabilities and robustness. It has been shown that, in comparison to regular DNN training, training with stylized images reduces texture biases in image classification and improves robustness with respect to image corruptions. In an effort to advance this line of research, we examine whether style transfer can likewise deliver these two effects in semantic segmentation. To this end, we perform style transfer with style varying across artificial image areas. Those random areas are formed by a chosen number of Voronoi cells. The resulting style-transferred data is then used to train semantic segmentation DNNs with the objective of reducing their dependence on texture cues while enhancing their reliance on shape-based features. In our experiments, it turns out that in semantic segmentation, style transfer augmentation reduces texture bias and strongly increases robustness with respect to common image corruptions as well as adversarial attacks. These observations hold for convolutional neural networks and transformer architectures on the Cityscapes dataset as well as on PASCAL Context, showing the generality of the proposed method.
Vanilla autoregressive image generation models generate visual tokens step-by-step, limiting their ability to capture holistic relationships among token sequences. Moreover, because most visual tokenizers map local image patches into latent tokens, global information is limited. To address this, we introduce \textit{Hita}, a novel image tokenizer for autoregressive (AR) image generation. It introduces a holistic-to-local tokenization scheme with learnable holistic queries and local patch tokens. Hita incorporates two key strategies to better align with the AR generation process: 1) {arranging} a sequential structure with holistic tokens at the beginning, followed by patch-level tokens, and using causal attention to maintain awareness of previous tokens; and 2) adopting a lightweight fusion module before feeding the de-quantized tokens into the decoder to control information flow and prioritize holistic tokens. Extensive experiments show that Hita accelerates the training speed of AR generators and outperforms those trained with vanilla tokenizers, achieving \textbf{2.59 FID} and \textbf{281.9 IS} on the ImageNet benchmark. Detailed analysis of the holistic representation highlights its ability to capture global image properties, such as textures, materials, and shapes. Additionally, Hita also demonstrates effectiveness in zero-shot style transfer and image in-painting. The code is available at \href{https://github.com/CVMI-Lab/Hita}{https://github.com/CVMI-Lab/Hita}.
Recent advancements in video generation, particularly in diffusion models, have driven notable progress in text-to-video (T2V) and image-to-video (I2V) synthesis. However, challenges remain in effectively integrating dynamic motion signals and flexible spatial constraints. Existing T2V methods typically rely on text prompts, which inherently lack precise control over the spatial layout of generated content. In contrast, I2V methods are limited by their dependence on real images, which restricts the editability of the synthesized content. Although some methods incorporate ControlNet to introduce image-based conditioning, they often lack explicit motion control and require computationally expensive training. To address these limitations, we propose AnyI2V, a training-free framework that animates any conditional images with user-defined motion trajectories. AnyI2V supports a broader range of modalities as the conditional image, including data types such as meshes and point clouds that are not supported by ControlNet, enabling more flexible and versatile video generation. Additionally, it supports mixed conditional inputs and enables style transfer and editing via LoRA and text prompts. Extensive experiments demonstrate that the proposed AnyI2V achieves superior performance and provides a new perspective in spatial- and motion-controlled video generation. Code is available at https://henghuiding.com/AnyI2V/.
Counterfactual explainability seeks to uncover model decisions by identifying minimal changes to the input that alter the predicted outcome. This task becomes particularly challenging for graph data due to preserving structural integrity and semantic meaning. Unlike prior approaches that rely on forward perturbation mechanisms, we introduce Graph Inverse Style Transfer (GIST), the first framework to re-imagine graph counterfactual generation as a backtracking process, leveraging spectral style transfer. By aligning the global structure with the original input spectrum and preserving local content faithfulness, GIST produces valid counterfactuals as interpolations between the input style and counterfactual content. Tested on 8 binary and multi-class graph classification benchmarks, GIST achieves a remarkable +7.6% improvement in the validity of produced counterfactuals and significant gains (+45.5%) in faithfully explaining the true class distribution. Additionally, GIST's backtracking mechanism effectively mitigates overshooting the underlying predictor's decision boundary, minimizing the spectral differences between the input and the counterfactuals. These results challenge traditional forward perturbation methods, offering a novel perspective that advances graph explainability.
Photorealistic style transfer (PST) enables real-world color grading by adapting reference image colors while preserving content structure. Existing methods mainly follow either approaches: generation-based methods that prioritize stylistic fidelity at the cost of content integrity and efficiency, or global color transformation methods such as LUT, which preserve structure but lack local adaptability. To bridge this gap, we propose Spatial Adaptive 4D Look-Up Table (SA-LUT), combining LUT efficiency with neural network adaptability. SA-LUT features: (1) a Style-guided 4D LUT Generator that extracts multi-scale features from the style image to predict a 4D LUT, and (2) a Context Generator using content-style cross-attention to produce a context map. This context map enables spatially-adaptive adjustments, allowing our 4D LUT to apply precise color transformations while preserving structural integrity. To establish a rigorous evaluation framework for photorealistic style transfer, we introduce PST50, the first benchmark specifically designed for PST assessment. Experiments demonstrate that SA-LUT substantially outperforms state-of-the-art methods, achieving a 66.7% reduction in LPIPS score compared to 3D LUT approaches, while maintaining real-time performance at 16 FPS for video stylization. Our code and benchmark are available at https://github.com/Ry3nG/SA-LUT
We introduce Color Disentangled Style Transfer (CDST), a novel and efficient two-stream style transfer training paradigm which completely isolates color from style and forces the style stream to be color-blinded. With one same model, CDST unlocks universal style transfer capabilities in a tuning-free manner during inference. Especially, the characteristics-preserved style transfer with style and content references is solved in the tuning-free way for the first time. CDST significantly improves the style similarity by multi-feature image embeddings compression and preserves strong editing capability via our new CDST style definition inspired by Diffusion UNet disentanglement law. By conducting thorough qualitative and quantitative experiments and human evaluations, we demonstrate that CDST achieves state-of-the-art results on various style transfer tasks.
We present a method for fine-grained control over music generation through inference-time interventions on an autoregressive generative music transformer called MusicGen. Our approach enables timbre transfer, style transfer, and genre fusion by steering the residual stream using weights of linear probes trained on it, or by steering the attention layer activations in a similar manner. We observe that modelling this as a regression task provides improved performance, hypothesizing that the mean-squared-error better preserve meaningful directional information in the activation space. Combined with the global conditioning offered by text prompts in MusicGen, our method provides both global and local control over music generation. Audio samples illustrating our method are available at our demo page.
While diffusion models have demonstrated remarkable generative capabilities, existing style transfer techniques often struggle to maintain identity while achieving high-quality stylization. This limitation is particularly acute for images where faces are small or exhibit significant camera-to-face distances, frequently leading to inadequate identity preservation. To address this, we introduce a novel, training-free framework for identity-preserved stylized image synthesis using diffusion models. Key contributions include: (1) the "Mosaic Restored Content Image" technique, significantly enhancing identity retention, especially in complex scenes; and (2) a training-free content consistency loss that enhances the preservation of fine-grained content details by directing more attention to the original image during stylization. Our experiments reveal that the proposed approach substantially surpasses the baseline model in concurrently maintaining high stylistic fidelity and robust identity integrity, particularly under conditions of small facial regions or significant camera-to-face distances, all without necessitating model retraining or fine-tuning.
Large Language Models (LLMs) and Multimodal LLMs have shown promising capabilities for SVG processing, yet existing benchmarks suffer from limited real-world coverage, lack of complexity stratification, and fragmented evaluation paradigms. We introduce SVGenius, a comprehensive benchmark comprising 2,377 queries across three progressive dimensions: understanding, editing, and generation. Built on real-world data from 24 application domains with systematic complexity stratification, SVGenius evaluates models through 8 task categories and 18 metrics. We assess 22 mainstream models spanning different scales, architectures, training paradigms, and accessibility levels. Our analysis reveals that while proprietary models significantly outperform open-source counterparts, all models exhibit systematic performance degradation with increasing complexity, indicating fundamental limitations in current approaches; however, reasoning-enhanced training proves more effective than pure scaling for overcoming these limitations, though style transfer remains the most challenging capability across all model types. SVGenius establishes the first systematic evaluation framework for SVG processing, providing crucial insights for developing more capable vector graphics models and advancing automated graphic design applications. Appendix and supplementary materials (including all data and code) are available at https://zju-real.github.io/SVGenius.
Expressive voice conversion aims to transfer both speaker identity and expressive attributes from a target speech to a given source speech. In this work, we improve over a self-supervised, non-autoregressive framework with a conditional variational autoencoder, focusing on reducing source timbre leakage and improving linguistic-acoustic disentanglement for better style transfer. To minimize style leakage, we use multilingual discrete speech units for content representation and reinforce embeddings with augmentation-based similarity loss and mix-style layer normalization. To enhance expressivity transfer, we incorporate local F0 information via cross-attention and extract style embeddings enriched with global pitch and energy features. Experiments show our model outperforms baselines in emotion and speaker similarity, demonstrating superior style adaptation and reduced source style leakage.
Zero-Shot Voice Conversion (VC) aims to transform the source speaker's timbre into an arbitrary unseen one while retaining speech content. Most prior work focuses on preserving the source's prosody, while fine-grained timbre information may leak through prosody, and transferring target prosody to synthesized speech is rarely studied. In light of this, we propose R-VC, a rhythm-controllable and efficient zero-shot voice conversion model. R-VC employs data perturbation techniques and discretize source speech into Hubert content tokens, eliminating much content-irrelevant information. By leveraging a Mask Generative Transformer for in-context duration modeling, our model adapts the linguistic content duration to the desired target speaking style, facilitating the transfer of the target speaker's rhythm. Furthermore, R-VC introduces a powerful Diffusion Transformer (DiT) with shortcut flow matching during training, conditioning the network not only on the current noise level but also on the desired step size, enabling high timbre similarity and quality speech generation in fewer sampling steps, even in just two, thus minimizing latency. Experimental results show that R-VC achieves comparable speaker similarity to state-of-the-art VC methods with a smaller dataset, and surpasses them in terms of speech naturalness, intelligibility and style transfer performance.
In medical image segmentation, limited external validity remains a critical obstacle when models are deployed across unseen datasets, an issue particularly pronounced in the ultrasound image domain. Existing solutions-such as domain adaptation and GAN-based style transfer-while promising, often fall short in the medical domain where datasets are typically small and diverse. This paper presents a novel application of principal component analysis (PCA) to address this limitation. PCA preprocessing reduces noise and emphasizes essential features by retaining approximately 90\% of the dataset variance. We evaluate our approach across six diverse breast tumor ultrasound datasets comprising 3,983 B-mode images and corresponding expert tumor segmentation masks. For each dataset, a corresponding dimensionality reduced PCA-dataset is created and U-Net-based segmentation models are trained on each of the twelve datasets. Each model trained on an original dataset was inferenced on the remaining five out-of-domain original datasets (baseline results), while each model trained on a PCA dataset was inferenced on five out-of-domain PCA datasets. Our experimental results indicate that using PCA reconstructed datasets, instead of original images, improves the model's recall and Dice scores, particularly for model-dataset pairs where baseline performance was lowest, achieving statistically significant gains in recall (0.57 $\pm$ 0.07 vs. 0.70 $\pm$ 0.05, $p = 0.0004$) and Dice scores (0.50 $\pm$ 0.06 vs. 0.58 $\pm$ 0.06, $p = 0.03$). Our method reduced the decline in recall values due to external validation by $33\%$. These findings underscore the potential of PCA reconstruction as a safeguard to mitigate declines in segmentation performance, especially in challenging cases, with implications for enhancing external validity in real-world medical applications.
Deep learning models often struggle to maintain performance when deployed on data distributions different from their training data, particularly in real-world applications where environmental conditions frequently change. While Multi-source Domain Generalization (MDG) has shown promise in addressing this challenge by leveraging multiple source domains during training, its practical application is limited by the significant costs and difficulties associated with creating multi-domain datasets. To address this limitation, we propose Pseudo Multi-source Domain Generalization (PMDG), a novel framework that enables the application of sophisticated MDG algorithms in more practical Single-source Domain Generalization (SDG) settings. PMDG generates multiple pseudo-domains from a single source domain through style transfer and data augmentation techniques, creating a synthetic multi-domain dataset that can be used with existing MDG algorithms. Through extensive experiments with PseudoDomainBed, our modified version of the DomainBed benchmark, we analyze the effectiveness of PMDG across multiple datasets and architectures. Our analysis reveals several key findings, including a positive correlation between MDG and PMDG performance and the potential of pseudo-domains to match or exceed actual multi-domain performance with sufficient data. These comprehensive empirical results provide valuable insights for future research in domain generalization. Our code is available at https://github.com/s-enmt/PseudoDomainBed.
CLIP is a discriminative model trained to align images and text in a shared embedding space. Due to its multimodal structure, it serves as the backbone of many generative pipelines, where a decoder is trained to map from the shared space back to images. In this work, we show that image synthesis is nevertheless possible using CLIP alone -- without any decoder, training, or fine-tuning. Our approach optimizes a frequency-aware implicit neural representation that encourages coarse-to-fine generation by stratifying frequencies across network layers. To stabilize this inverse mapping, we introduce adversarially robust initialization, a lightweight Orthogonal Procrustes projection to align local text and image embeddings, and a blending loss that anchors outputs to natural image statistics. Without altering CLIP's weights, this framework unlocks capabilities such as text-to-image generation, style transfer, and image reconstruction. These findings suggest that discriminative models may hold untapped generative potential, hidden in plain sight.
Gaussian Splatting (GS) has recently emerged as an efficient representation for rendering 3D scenes from 2D images and has been extended to images, videos, and dynamic 4D content. However, applying style transfer to GS-based representations, especially beyond simple color changes, remains challenging. In this work, we introduce CLIPGaussians, the first unified style transfer framework that supports text- and image-guided stylization across multiple modalities: 2D images, videos, 3D objects, and 4D scenes. Our method operates directly on Gaussian primitives and integrates into existing GS pipelines as a plug-in module, without requiring large generative models or retraining from scratch. CLIPGaussians approach enables joint optimization of color and geometry in 3D and 4D settings, and achieves temporal coherence in videos, while preserving a model size. We demonstrate superior style fidelity and consistency across all tasks, validating CLIPGaussians as a universal and efficient solution for multimodal style transfer.
This work introduces Ui2i, a novel model for unpaired image-to-image translation, trained on content-wise unpaired datasets to enable style transfer across domains while preserving content. Building on CycleGAN, Ui2i incorporates key modifications to better disentangle content and style features, and preserve content integrity. Specifically, Ui2i employs U-Net-based generators with skip connections to propagate localized shallow features deep into the generator. Ui2i removes feature-based normalization layers from all modules and replaces them with approximate bidirectional spectral normalization -- a parameter-based alternative that enhances training stability. To further support content preservation, channel and spatial attention mechanisms are integrated into the generators. Training is facilitated through image scale augmentation. Evaluation on two biomedical tasks -- domain adaptation for nuclear segmentation in immunohistochemistry (IHC) images and unmixing of biological structures superimposed in single-channel immunofluorescence (IF) images -- demonstrates Ui2i's ability to preserve content fidelity in settings that demand more accurate structural preservation than typical translation tasks. To the best of our knowledge, Ui2i is the first approach capable of separating superimposed signals in IF images using real, unpaired training data.
Recent advances in expressive text-to-speech (TTS) have introduced diverse methods based on style embedding extracted from reference speech. However, synthesizing high-quality expressive speech remains challenging. We propose Spotlight-TTS, which exclusively emphasizes style via voiced-aware style extraction and style direction adjustment. Voiced-aware style extraction focuses on voiced regions highly related to style while maintaining continuity across different speech regions to improve expressiveness. We adjust the direction of the extracted style for optimal integration into the TTS model, which improves speech quality. Experimental results demonstrate that Spotlight-TTS achieves superior performance compared to baseline models in terms of expressiveness, overall speech quality, and style transfer capability. Our audio samples are publicly available.
Given a pair of source and reference speech recordings, audio-to-audio (A2A) style transfer involves the generation of an output speech that mimics the style characteristics of the reference while preserving the content and speaker attributes of the source. In this paper, we propose a novel framework, termed as A2A Zero-shot Emotion Style Transfer (A2A-ZEST), that enables the transfer of reference emotional attributes to the source while retaining its speaker and speech contents. The A2A-ZEST framework consists of an analysis-synthesis pipeline, where the analysis module decomposes speech into semantic tokens, speaker representations, and emotion embeddings. Using these representations, a pitch contour estimator and a duration predictor are learned. Further, a synthesis module is designed to generate speech based on the input representations and the derived factors. This entire paradigm of analysis-synthesis is trained purely in a self-supervised manner with an auto-encoding loss. For A2A emotion style transfer, the emotion embedding extracted from the reference speech along with the rest of the representations from the source speech are used in the synthesis module to generate the style translated speech. In our experiments, we evaluate the converted speech on content/speaker preservation (w.r.t. source) as well as on the effectiveness of the emotion style transfer (w.r.t. reference). The proposal, A2A-ZEST, is shown to improve over other prior works on these evaluations, thereby enabling style transfer without any parallel training data. We also illustrate the application of the proposed work for data augmentation in emotion recognition tasks.
During the finetuning stage of text generation tasks, standard cross-entropy loss treats all tokens equally. This can lead models to overemphasize high-frequency, low-information tokens, neglecting lower-frequency tokens crucial for specificity and informativeness in generated content. This paper introduces a novel loss function, Power-Law Decay Loss (PDL), specifically designed to optimize the finetuning process for text generation. The core motivation for PDL stems from observations in information theory and linguistics: the informativeness of a token is often inversely proportional to its frequency of occurrence. PDL re-weights the contribution of each token in the standard cross-entropy loss based on its frequency in the training corpus, following a power-law decay. Specifically, the weights for high-frequency tokens are reduced, while low-frequency, information-dense tokens are assigned higher weights. This mechanism guides the model during finetuning to focus more on learning and generating tokens that convey specific and unique information, thereby enhancing the quality, diversity, and informativeness of the generated text. We theoretically elaborate on the motivation and construction of PDL and discuss its potential applications and advantages across various text generation finetuning tasks, such as abstractive summarization, dialogue systems, and style transfer.
Semantic segmentation models trained on synthetic data often perform poorly on real-world images due to domain gaps, particularly in adverse conditions where labeled data is scarce. Yet, recent foundation models enable to generate realistic images without any training. This paper proposes to leverage such diffusion models to improve the performance of vision models when learned on synthetic data. We introduce two novel techniques for semantically consistent style transfer using diffusion models: Class-wise Adaptive Instance Normalization and Cross-Attention (CACTI) and its extension with selective attention Filtering (CACTIF). CACTI applies statistical normalization selectively based on semantic classes, while CACTIF further filters cross-attention maps based on feature similarity, preventing artifacts in regions with weak cross-attention correspondences. Our methods transfer style characteristics while preserving semantic boundaries and structural coherence, unlike approaches that apply global transformations or generate content without constraints. Experiments using GTA5 as source and Cityscapes/ACDC as target domains show that our approach produces higher quality images with lower FID scores and better content preservation. Our work demonstrates that class-aware diffusion-based style transfer effectively bridges the synthetic-to-real domain gap even with minimal target domain data, advancing robust perception systems for challenging real-world applications. The source code is available at: https://github.com/echigot/cactif.