arXiv:2511.21029v1 Announce Type: new 
Abstract: Music-to-dance generation aims to translate auditory signals into expressive human motion, with broad applications in virtual reality, choreography, and digital entertainment. Despite promising progress, the limited generation efficiency of existing methods leaves insufficient computational headroom for high-fidelity 3D rendering, thereby constraining the expressiveness of 3D characters during real-world applications. Thus, we propose FlowerDance, which not only generates refined motion with physical plausibility and artistic expressiveness, but also achieves significant generation efficiency on inference speed and memory utilization . Specifically, FlowerDance combines MeanFlow with Physical Consistency Constraints, which enables high-quality motion generation with only a few sampling steps. Moreover, FlowerDance leverages a simple but efficient model architecture with BiMamba-based backbone and Channel-Level Cross-Modal Fusion, which generates dance with efficient non-autoregressive manner. Meanwhile, FlowerDance supports motion editing, enabling users to interactively refine dance sequences. Extensive experiments on AIST++ and FineDance show that FlowerDance achieves state-of-the-art results in both motion quality and generation efficiency. Code will be released upon acceptance.

تقدم FlowerDance نهجًا جديدًا لتوليد الرقص من الموسيقى، باستخدام MeanFlow لإنشاء حركات رقص ثلاثية الأبعاد مصقولة وقابلة للتطبيق جسديًا بكفاءة. تعزز هذه الطريقة تعبير الشخصيات ثلاثية الأبعاد بينما تعالج القيود الحاسوبية للتقنيات الحالية، مما يسمح بتقديم عالي الدقة في التطبيقات الواقعية.

FlowerDance presenta un enfoque novedoso para la generación de danza a partir de música, utilizando MeanFlow para crear movimientos de danza 3D refinados y físicamente plausibles de manera eficiente. Este método mejora la expresividad de los personajes 3D mientras aborda las limitaciones computacionales de las técnicas existentes, permitiendo un renderizado de alta fidelidad en aplicaciones del mundo real.

FlowerDance propose une nouvelle approche pour la génération de danse à partir de musique, utilisant MeanFlow pour créer des mouvements de danse 3D raffinés et physiquement plausibles de manière efficace. Cette méthode améliore l'expressivité des personnages 3D tout en répondant aux limitations computationnelles des techniques existantes, permettant un rendu haute fidélité dans des applications réelles.

FlowerDance introduces a novel approach to music-to-dance generation, utilizing MeanFlow to create refined and physically plausible 3D dance motions efficiently. This method enhances the expressiveness of 3D characters while addressing the computational limitations of existing techniques, allowing for high-fidelity rendering in real-world applications.

FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation

arXiv:2511.21215v1 Announce Type: new 
Abstract: We present a comprehensive comparative study of three generative modeling paradigms: Denoising Diffusion Probabilistic Models (DDPM), Conditional Flow Matching (CFM), and MeanFlow. While DDPM and CFM require iterative sampling, MeanFlow enables direct one-step generation by modeling the average velocity over time intervals. We implement all three methods using a unified TinyUNet architecture (<1.5M parameters) on CIFAR-10, demonstrating that CFM achieves an FID of 24.15 with 50 steps, significantly outperforming DDPM (FID 402.98). MeanFlow achieves FID 29.15 with single-step sampling -- a 50X reduction in inference time. We further extend CFM to image inpainting, implementing mask-guided sampling with four mask types (center, random bbox, irregular, half). Our fine-tuned inpainting model achieves substantial improvements: PSNR increases from 4.95 to 8.57 dB on center masks (+73%), and SSIM improves from 0.289 to 0.418 (+45%), demonstrating the effectiveness of inpainting-aware training.

تم إجراء دراسة مقارنة شاملة تقارن بين ثلاثة نماذج للتوليد: نماذج التوزيع الاحتمالي لإزالة الضوضاء (DDPM) ، ومطابقة التدفق الشرطي (CFM) ، وMeanFlow ، مع التركيز على تطبيقها في إعادة تشكيل الصور. تسلط الدراسة الضوء على أن CFM يتفوق بشكل كبير على DDPM من حيث الكفاءة والجودة ، حيث حقق درجة FID ملحوظة قدرها 24.15 مع 50 خطوة فقط ، بينما يسمح MeanFlow بالتوليد في خطوة واحدة ، مما يقلل من وقت الاستدلال بمقدار 50 مرة.

Se ha llevado a cabo un estudio comparativo exhaustivo que analiza tres paradigmas de modelado generativo: Modelos Probabilísticos de Difusión de Desenfoque (DDPM), Coincidencia de Flujo Condicional (CFM) y MeanFlow, centrándose en su aplicación en la restauración de imágenes. El estudio destaca que el CFM supera significativamente al DDPM en términos de eficiencia y calidad, logrando un notable puntaje FID de 24.15 con solo 50 pasos, mientras que MeanFlow permite la generación en un solo paso, reduciendo el tiempo de inferencia en 50 veces.

Une étude comparative approfondie a été réalisée sur trois paradigmes de modélisation générative : les modèles probabilistes de diffusion de débruitage (DDPM), l'appariement de flux conditionnel (CFM) et MeanFlow, en se concentrant sur leur application dans le remplissage d'images. L'étude souligne que le CFM surpasse significativement le DDPM en termes d'efficacité et de qualité, atteignant un score FID remarquable de 24,15 avec seulement 50 étapes, tandis que MeanFlow permet une génération en une seule étape, réduisant le temps d'inférence par 50 fois.

A comprehensive study has been conducted comparing three generative modeling paradigms: Denoising Diffusion Probabilistic Models (DDPM), Conditional Flow Matching (CFM), and MeanFlow, focusing on their application in image inpainting. The study highlights that CFM significantly outperforms DDPM in terms of efficiency and quality, achieving a notable FID score of 24.15 with only 50 steps, while MeanFlow allows for single-step generation, reducing inference time by 50 times.

From Diffusion to One-Step Generation: A Comparative Study of Flow-Based Models with Application to Image Inpainting

arXiv:2511.19065v1 Announce Type: cross 
Abstract: MeanFlow promises high-quality generative modeling in few steps, by jointly learning instantaneous and average velocity fields. Yet, the underlying training dynamics remain unclear. We analyze the interaction between the two velocities and find: (i) well-established instantaneous velocity is a prerequisite for learning average velocity; (ii) learning of instantaneous velocity benefits from average velocity when the temporal gap is small, but degrades as the gap increases; and (iii) task-affinity analysis indicates that smooth learning of large-gap average velocities, essential for one-step generation, depends on the prior formation of accurate instantaneous and small-gap average velocities. Guided by these observations, we design an effective training scheme that accelerates the formation of instantaneous velocity, then shifts emphasis from short- to long-interval average velocity. Our enhanced MeanFlow training yields faster convergence and significantly better few-step generation: With the same DiT-XL backbone, our method reaches an impressive FID of 2.87 on 1-NFE ImageNet 256x256, compared to 3.43 for the conventional MeanFlow baseline. Alternatively, our method matches the performance of the MeanFlow baseline with 2.5x shorter training time, or with a smaller DiT-L backbone.

أدت التطورات الأخيرة في تدريب MeanFlow إلى توضيح الديناميات بين مجالات السرعة الفورية والمتوسطة، مما يكشف أن التعلم الفعال للسرعة المتوسطة يعتمد على التأسيس المسبق لسرعات فورية دقيقة. وقد أدى هذا الفهم إلى تصميم مخطط تدريب جديد يسرع من تشكيل هذه السرعات، مما يحسن العملية التدريبية بشكل عام.

Los recientes avances en el entrenamiento de MeanFlow han aclarado la dinámica entre los campos de velocidad instantánea y promedio, revelando que el aprendizaje efectivo de la velocidad promedio depende del establecimiento previo de velocidades instantáneas precisas. Esta comprensión ha llevado al diseño de un nuevo esquema de entrenamiento que acelera la formación de estas velocidades, mejorando así el proceso de entrenamiento en general.

Les récentes avancées dans l'entraînement de MeanFlow ont clarifié la dynamique entre les champs de vitesse instantanée et moyenne, révélant que l'apprentissage efficace de la vitesse moyenne dépend de l'établissement préalable de vitesses instantanées précises. Cette compréhension a conduit à la conception d'un nouveau schéma d'entraînement qui accélère la formation de ces vitesses, améliorant ainsi le processus d'entraînement global.

Recent advancements in MeanFlow training have clarified the dynamics between instantaneous and average velocity fields, revealing that effective learning of average velocity relies on the prior establishment of accurate instantaneous velocities. This understanding has led to the design of a new training scheme that accelerates the formation of these velocities, enhancing the overall training process.

FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation

Was this article worth reading? Share it

Fakeface

Rendora AI

FLUX AI ART