High-fidelity hand gesture generation represents a significant challenge in human-centric generation tasks. Existing methods typically employ a single-view mesh-rendered image prior to enhancing gesture generation quality. However, the spatial complexity of hand gestures and the inherent limitations of single-view rendering make it difficult to capture complete gesture information, particularly when fingers are occluded. The fundamental contradiction lies in the loss of 3D topological relationships through 2D projection and the incomplete spatial coverage inherent to single-view representations. Diverging from single-view prior approaches, we propose a multi-view prior framework, named Multi-Modal UNet-based Feature Encoder (MUFEN), to guide diffusion models in learning comprehensive 3D hand information. Specifically, we extend conventional front-view rendering to include rear, left, right, top, and bottom perspectives, selecting the most information-rich view combination as training priors to address occlusion. This multi-view prior with a dedicated dual stream encoder significantly improves the model's understanding of complete hand features. Furthermore, we design a bounding box feature fusion module, which can fuse the gesture localization features and multi-modal features to enhance the location-awareness of the MUFEN features to the gesture-related features. Experiments demonstrate that our method achieves state-of-the-art performance on quantitative metrics and exhibits superior qualitative results. The source code is available at https://github.com/fuqifan/MUFEN.
论文检索
输入标题、作者或关键词，从 9,256 篇学术成果中精准定位
Recently, diffusion models have demonstrated powerful capabilities in image generation. However, the repetitive and sequential denoising process adds significant time and computational costs, limiting their application. In this paper, we propose a training-free and universal method, Horizontal-Vertical Accelerated Denoising (HVAD). It mainly utilizes the inherent temporal redundancy in the diffusion model to parallelize the time steps of the denoising process horizontally and reduces the computation of individual denoising time steps vertically. It achieves simultaneous inference acceleration from both the horizontal and vertical directions of the denoising process. Thus, significant inference acceleration is achieved without sacrificing generation quality. Experimental results based on the MS-COCO validation set show that the method achieves a practical speedup of 1.97× on Stable Diffusion v1.5, and achieves a theoretical speedup of 2.40× or even higher with guaranteed generation quality. To verify the generality of the method, better results were achieved on other versions of Stable Diffusion, such as Stable Diffusion v2.1 and Stable Diffusion XL.
Despite significant progress in diffusion-based image generation, subject-driven generation and instruction-based editing remain challenging. Existing methods typically treat them separately, struggling with limited high-quality data and poor generalization. However, both tasks require capturing complex visual variations while maintaining consistency between inputs and outputs. Inspired by this, we propose MIGE, a unified framework that standardizes task representations using multimodal instructions. It first treats subject-driven generation as creation on a blank canvas and instruction-based editing as modification of an existing image, establishing a shared input-output formulation, then introduces a novel multimodal encoder that maps free-form multimodal instructions into a unified vision-language space, integrating visual and semantic features through a feature fusion mechanism. This unification enables joint training of both tasks, providing two key advantages: (1) Cross-Task Enhancement: by leveraging shared visual and semantic representations, joint training improves instruction adherence and visual consistency in both subject-driven generation and instruction-based editing. (2) Generalization: learning in a unified format facilitates cross-task knowledge transfer, enabling MIGE to generalize to novel compositional tasks, including instruction-based subject-driven editing. Experiments show that MIGE excels in both subject-driven generation and instruction-based editing while setting a SOTA in the new task of instruction-based subject-driven editing. Code and model have been publicly available at https://github.com/Eureka-Maggie/MIGE.
Cloth-changing Person Re-Identification (CCReID) aims to recognize individuals across clothing variations by learning clothing-invariant representations. However, obtaining sufficient samples of the same person in diverse outfits is often impractical. While synthesizing realistic person images provides an effective solution, existing augmentation methods require labeled data and external priors (e.g., pose skeletons, semantic maps), resulting in high costs and limited generalization. To this end, we propose a Prior-Free Augmentation method for Cloth-changing person re-identification (PFAC), which leverages text guidance to synthesize images with clothing variations while maintaining identity consistency. Our approach features: (1) a truncated diffusion model that preserves clothing-invariant structural cues from intermediate noisy images, (2) a dual-branch denoising network that decouples text-guided clothing synthesis from identity consistency via cross-modal alignment, and (3) a joint optimization strategy with identity-focused losses and image filtering to enhance realism and discriminability. Experimental results on PRCC, LTCC, and Celeb-reID datasets demonstrate that PFAC achieves state-of-the-art CCReID performance, effectively generating high-fidelity, identity-consistent images for robust augmentation without external priors.
Hairstyles are intricate and culturally significant with various geometries, textures, and structures. Existing text or image-guided generation methods fail to handle the richness and complexity of diverse styles. We present a novel approach for 3D hair strand generation that accommodates diverse image inputs across styles, viewpoints, and quantities of input views. Our method consists of three key components. First, our MultiHair Dataset provides 457 diverse hairstyles annotated with 74 attributes, emphasizing complex and culturally significant styles to improve model generalization. Second, we propose a diffusion framework conditioned on multi-view linearts that can capture topological cues (e.g., strand density and parting lines) while filtering out noise. By leveraging a latent diffusion model with cross-attention on lineart features, our method achieves flexible and robust 3D hair generation across diverse input conditions. Third, a parametric post-processing module enforces braid-specific constraints to maintain coherence in complex structures. This framework not only advances hairstyle realism and diversity but also enables culturally inclusive digital avatars and novel applications like sketch-based 3D strand editing for animation and augmented reality.
High-quality three-dimensional (3D) reconstruction from sparse views is critical for applications such as virtual and augmented reality, robotics, and digital content creation. While methods like Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have shown strong performance in novel view synthesis, they struggle in few-shot settings, especially when scenes contain large occluded or unseen regions. The lack of explicit supervision for hidden content limits reconstruction completeness and realism. We propose See-Through-the-Occlusion Gaussian Splatting (STO-GS), a novel framework that rethinks occlusion modeling in static scenes. Drawing inspiration from four-dimensional Gaussian Splatting (4DGS), we reinterpret time as a proxy for occlusion depth and apply deformation-based opacity modulation to recover hidden layers. To provide supervision, we generate amodal views via diffusion-based inpainting, exposing occluded structures for training. A two-stage layered training pipeline further refines the reconstruction, with a multi-layer perceptron (MLP) adjusting Gaussian opacity across occlusion layers. STO-GS improves occlusion-aware reconstruction and achieves superior performance over existing few-shot 3DGS baselines, including a 0.51 dB gain on challenging datasets.
Fine-tuning large-scale music audio generation models, such as MusicGen and Mustango, is a computationally expensive process, often requiring updates to billions of parameters and, therefore, significant hardware resources. Parameter-Efficient Fine-Tuning (PEFT) techniques, particularly adapter-based methods, have emerged as a promising alternative, enabling adaptation with minimal trainable parameters while preserving model performance. However, the design choices for adapters, including their architecture, placement, and size, are numerous, and it is unclear which of these combinations would produce optimal adapters and why, for a given case of low-resource music genre. In this paper, we attempt to answer this question by studying various adapter configurations for two AI music models, MusicGen and Mustango, on two genres: Hindustani Classical and Turkish Makam music. Our findings reveal distinct trade-offs: convolution-based adapters excel in capturing fine-grained local musical details such as ornamentations and short melodic phrases, while transformer-based adapters better preserve long-range dependencies crucial for structured improvisation. Additionally, we analyze computational resource requirements across different adapter scales, demonstrating how mid-sized adapters (40M parameters) achieve an optimal balance between expressivity and quality. Furthermore, we find that Mustango, a diffusion-based model, generates more diverse outputs with better adherence to the description in the input prompt while lacking in providing stability in notes, rhythm alignment, and aesthetics. Also, it is computationally intensive and requires significantly more time to train. In contrast, autoregressive models like MusicGen offer faster training and are more efficient, and can produce better quality output in comparison, but have slightly higher redundancy in their generations. We release our datasets, models and training code in the following github repository: https://github.com/atharva20038/ACMMM_Adapters/tree/main.
The success of diffusion models has enabled effortless, high-quality image modifications that precisely align with users' intentions, thereby raising concerns about their potential misuse by malicious actors. Previous studies have attempted to mitigate such misuse through adversarial attacks. However, these approaches heavily rely on image-level inconsistencies, which pose fundamental limitations in addressing the influence of textual prompts. In this paper, we propose PromptFlare, a novel adversarial protection method designed to protect images from malicious modifications facilitated by diffusion-based inpainting models. Our approach leverages the cross-attention mechanism to exploit the intrinsic properties of prompt embeddings. Specifically, we identify and target shared token of prompts that is invariant and semantically uninformative, injecting adversarial noise to suppress the sampling process. The injected noise acts as a cross-attention decoy, diverting the model's focus away from meaningful prompt-image alignments and thereby neutralizing the effect of prompt. Extensive experiments on the EditBench dataset demonstrate that our method achieves state-of-the-art performance across various metrics while significantly reducing computational overhead and GPU memory usage. These findings highlight PromptFlare as a robust and efficient protection against unauthorized image manipulations. The code is available at https://github.com/NAHOHYUN-SKKU/PromptFlare.
Diffusion-based text-to-image models have demonstrated remarkable capabilities in generating realistic images, but they raise societal and ethical concerns, such as the creation of unsafe content. While concept editing is proposed to address these issues, they often struggle to balance the removal of unsafe concept with maintaining the model's general generative capabilities. In this work, we propose ACE, a new editing method that enhances concept editing in diffusion models. ACE introduces a novel cross null-space projection approach to precisely erase unsafe concept while maintaining the model's ability to generate high-quality, semantically consistent images. Extensive experiments demonstrate that ACE significantly outperforms the advancing baselines, improving semantic consistency by 24.56% and image generation quality by 34.82% on average with only 1% of the time cost. These results highlight the practical utility of concept editing by mitigating its potential risks, paving the way for broader applications in the field. WARNING: This paper contains harmful content that can be offensive.
Recent Diffusion Transformers (DiTs) have shown impressive capabilities in generating single-modality content, including images, videos, and audio. However, the potential of DiTs to enable superb multimodal content creation remains underexplored. To bridge this gap, we introduce AV-DiT, a novel and efficient audio-visual diffusion transformer designed to generate high-quality, realistic videos with synchronized audio tracks. To minimize model complexity and computational costs, our AV-DiT utilizes a modality-shared DiT backbone pre-trained on image-only data, with only newly inserted adapters being trainable. This shared backbone facilitates the generation of both audio and video. Specifically, the video branch incorporates a trainable temporal attention layer into a pre-trained DiT block for capturing the temporal consistency for video generation. In addition, a small number of trainable parameters adapt the image-based DiT block to learn the acoustic characteristics for audio generation. An extra shared self-attention block reused from the DiT block, equipped with lightweight parameters, facilitates feature interaction between audio and visual modalities for alignment. Extensive experiments on the datasets demonstrate that our AV-DiT achieves state-of-the-art performance in joint audio-visual generation with significantly fewer tunable parameters. Furthermore, our results highlight that a single shared image generative backbone with modality-specific adaptations is sufficient for constructing a joint audio-video generator.
Audio-driven video generation aims to synthesize realistic videos that align with input audio recordings, akin to the human ability to visualize scenes from auditory input. However, existing approaches predominantly focus on exploring semantic information, such as the classes of sounding sources present in the audio, limiting their ability to generate videos with accurate content and spatial composition. In contrast, we humans can not only naturally identify the semantic categories of sounding sources but also determine their deeply encoded spatial attributes, including locations and movement directions. This useful information can be elucidated by considering specific spatial indicators derived from the inherent physical properties of sound, such as loudness or frequency. As prior methods largely ignore this factor, we present SpA2V, the first framework explicitly exploits these spatial auditory cues from audios to generate videos with high semantic and spatial correspondence. SpA2V decomposes the generation process into two stages: 1) Audio-guided Video Planning: We meticulously adapt a state-of-the-art MLLM for a novel task of harnessing spatial and semantic cues from input audio to construct Video Scene Layouts (VSLs). This serves as an intermediate representation to bridge the gap between the audio and video modalities. 2) Layout-grounded Video Generation: We develop an efficient and effective approach to seamlessly integrate VSLs as conditional guidance into pre-trained diffusion models, enabling VSL-grounded video generation in a training-free manner. Extensive experiments demonstrate that SpA2V excels in generating realistic videos with semantic and spatial alignment to the input audios.
Multi-conditional image generation aims to create customized images that align with multiple specified conditions. Existing methods, whether through end-to-end training or by fine-tuning adapters to integrate pre-trained control modules of the same category (e.g., LoRA, IP-Adapter, ControlNet, T2I-Adapter), are restricted to a closed set of predefined input conditions. To overcome this limitation, we propose ModuleTeam, a training-free method for latent mixture of arbitrary control modules, capable of handling open-set conditions by incorporating the corresponding modules. The design of ModuleTeam is rooted in two key findings: (i) modules interfere with each other at the level of model parameters, and (ii) module weights contribute to the generated images by affecting the noise predictions within the diffusion process in an approximately linear manner. The first finding motivates our latent mixture approach, which mixes the control modules by aggregating their latent variables between diffusion model blocks. The second finding enables a multi-inference module reweighting strategy that balances module contributions to generation, requiring no additional training or fine-tuning overhead. Extensive results demonstrate that ModuleTeam not only outperforms existing methods but also provides flexibility in the types of conditions and scalability in their number.
Image editing requires semantically modifying specific regions according to user instructions while preserving overall visual coherence. Although diffusion models have shown remarkable progress in image generation, their application to editing tasks faces two critical limitations: (1) insufficient understanding of editing objectives often leads to inconsistencies in style, attribute, or texture between generated content and background regions, and (2) over-reliance on ambiguous textual prompts that frequently lack crucial details, resulting in suboptimal edits. To address these challenges, we propose SAKR-Edit, a novel framework that enhances editing quality and controllability through Scene-Aware Knowledge Reasoning. Specifically, our approach introduces a scene-aware knowledge reasoning module that combines large language models (LLMs) with vision-language models (e.g., BLIP-2) to integrate global and local semantic information for improved instruction comprehension. The system employs chain-of-thought reasoning and contextual learning to parse instructions, infer implicit editing intentions, and supplement missing details, thereby improving editing precision. Additionally, we construct SSUD, a structured scene understanding dataset for evaluating editing models in real-world scenarios. Extensive experiments demonstrate that SAKR-Edit outperforms existing methods in image realism, style consistency, and structural integrity, while showing robust stability and adaptability in real-world applications. Our code and dataset are released at https://github.com/SAKR-Edit/sakr-edit.github.io.
Dance is an important art form in human culture, but creating new dances can be both challenging and time-consuming. In this paper, we propose a novel dance choreography framework, EDMG, designed to efficiently generate creative and long-lasting dance sequences conditioning on music and dance descriptions. In the first stage, we propose a flexible dance diffusion method, combined with dance genre description and descriptions of fundamental movements to generate the dance sequences. To achieve high computational efficiency and inference speed, EDMG designs a lightweight denoising module by using selective parallel scanning algorithm from Mamba2. This Parallel Mamba Denoiser reduces significantly the number of parameters and accelerates remarkably both the learning and inference processes. In the second stage, by designing a smoothing module with a long receptive field, we mitigate joint error accumulation that causes jittering movements and foot sliding, thereby enhancing the fluency and visual appeal of the dance movements. Furthermore, we extend the AIST++ dataset by adding detailed descriptions of dance genres and fundamental movements, using the Large Language Model (LLM). These descriptions further improve the choreography generation. EDMG is validated through extensive experiments, demonstrating that our method can both effectively and efficiently generate long-term dances suitable for various dance genres. Project URL: https://github.com/neymar277/EDMG.
Diffusion-based video generation techniques have significantly improved zero-shot talking-head avatar generation, enhancing the naturalness of both head motion and facial expressions. However, existing methods suffer from poor controllability, making them less applicable to real-world scenarios such as filmmaking and live streaming for e-commerce. To address this limitation, we propose FLAP, a novel approach that integrates explicit 3D intermediate parameters (head poses and facial expressions) into the diffusion model for end-to-end generation of realistic portrait videos. The proposed architecture allows the model to generate vivid portrait videos from audio while simultaneously incorporating additional control signals, such as head rotation angles and eye-blinking frequency. Furthermore, the decoupling of head pose and facial expression allows for independent control of each, offering precise manipulation of both the avatar's pose and facial expressions. We also demonstrate its flexibility in integrating with existing 3D head generation methods, bridging the gap between 3D model-based approaches and end-to-end diffusion techniques. Extensive experiments show that our method outperforms recent audio-driven portrait video models in both naturalness and controllability.
Diffusion models have demonstrated remarkable success in generative tasks, yet their iterative denoising process results in slow inference, limiting their practicality. While existing acceleration methods exploit the well-known U-shaped similarity pattern between adjacent steps through caching mechanisms, they lack theoretical foundation and rely on simplistic computation reuse, often leading to performance degradation. In this work, we provide a theoretical understanding by analyzing the denoising process through the second-order Adams-Bashforth method, revealing a linear relationship between the outputs of consecutive steps. This analysis explains why the outputs of adjacent steps exhibit a U-shaped pattern. Furthermore, extending Adams-Bashforth method to higher order, we propose a novel caching-based acceleration approach for diffusion models, instead of directly reusing cached results, with a truncation error bound of only (O(hk) where h is the step size. Extensive validation across diverse image and video diffusion models (including HunyuanVideo and FLUX.1-dev) with various schedulers demonstrates our method's effectiveness in achieving nearly 3× speedup while maintaining original performance levels, offering a practical real-time solution without compromising generation quality.
Diffusion models exhibit impressive generative capabilities but are significantly impacted by exposure bias. In this paper, we make a key observation: the energy of predicted noisy samples in the reverse process continuously declines compared to perturbed samples in the forward process. Building on this, we identify two important findings: 1) The reduction in energy follows distinct patterns in the low-frequency and high-frequency subbands; 2) The subband energy of reverse-process reconstructed samples is consistently lower than that of forward-process ones, and both are lower than the original data samples. Based on the first finding, we introduce a dynamic frequency regulation mechanism utilizing wavelet transforms, which separately adjusts the low- and high-frequency subbands. Leveraging the second insight, we derive the rigorous mathematical form of exposure bias. It is worth noting that, our method is training-free and plug-and-play, significantly improving the generative quality of various diffusion models and frameworks with negligible computational cost. The source code is available at https://github.com/kunzhan/wpp.
Text-to-image diffusion models have made significant progress in image generation, allowing for effortless customized generation. However, existing image editing methods still face certain limitations when dealing with personalized image composition tasks. First, there is the issue of lack of geometric control over the inserted objects. Current methods are confined to 2D space and typically rely on textual instructions, making it challenging to maintain precise geometric control over the objects. Second, there is the challenge of style consistency. Existing methods often overlook the style consistency between the inserted object and the background, resulting in a lack of realism. In addition, the challenge of inserting objects into images without extensive training remains significant. To address these issues, we propose FreeInsert, a novel training-free framework that customizes object insertion into arbitrary scenes by leveraging 3D geometric information. Benefiting from the advances in existing 3D generation models, we first convert the 2D object into 3D, perform interactive editing at the 3D level, and then re-render it into a 2D image from a specified view. This process introduces geometric controls such as shape or view. The rendered image, serving as geometric control, is combined with style and content control achieved through diffusion adapters, ultimately producing geometrically controlled, style-consistent edited images via the diffusion model.
Dataset condensation distills a large dataset into a small synthetic surrogate dataset with similar training efficacy on downstream tasks. Of the existing condensation methods, diffusion-based methods that synthesize surrogate datasets with diffusion models have successfully distilled high-resolution datasets with high training efficacy and satisfactory cross-architectural transferability. However, these methods exhibit a random sampling bias that impairs their performance in dataset condensation settings. We propose a novel dataset condensation method called Noise-Optimized Distribution Distillation (NODD) that mitigates this sampling bias to improve the training performance of synthetic datasets generated with diffusion models. NODD can integrate with existing diffusion-based methods to produce synthetic datasets with enhanced training performance.
Recent diffusion model advancements aim to handle conditional generative tasks without extra training. Existing training-free methods add a correction term at each denoising step, but they often face computational instability and lack controllability, especially with limited samples and large noise. We propose a new approach using the von Mises-Fisher (vMF) distribution to model the denoised result, turning the conditional generation task into an estimation problem for vMF parameters. We formulate the conditional diffusion model as a mean vector estimation problem for the Gaussian distribution, noting that this can be seen as an estimation problem from noisy observations. When the sampling number is small, the estimation is unstable. To address this, we optimize the mean vector of the vMF distribution by minimizing the KL divergence between the prior and posterior distributions. This approach not only addresses the computational instability but also improves the controllability and quality of the generated results. Once these parameters are determined, the denoised result can be sampled directly from the vMF distribution. Estimating the parameters requires minimal additional code and incurs negligible computational overhead while significantly improving performance. Extensive experiments across various conditional generation tasks, including depth maps, edge detection, segmentation, and style guidance, demonstrate the superiority and versatility of our method. Our approach consistently outperforms existing training-free methods and even surpasses some training-required methods in terms of visual quality and controllability.