Avatar V: Scaling Video-Reference Avatar Video Generation

A new production-scale framework called Avatar V can generate avatar videos that reproduce a person’s talking rhythm, gestural tendencies, and expression dynamics by conditioning directly on the full token sequence of a reference video, according to research published on arXiv [1]. The system departs from earlier methods that rely on single static images, which the authors say provide insufficient identity information and cannot capture dynamic motion traits [1]. Instead, Avatar V learns both static identity attributes — such as facial geometry and skin texture — and dynamic behavioral patterns including micro-expressions through attention over the reference context [1]. The framework introduces Sparse Reference Attention, an asymmetric mechanism that achieves linear-complexity conditioning on arbitrarily long references, along with a motion representation stream for closed-loop talking style transfer and an identity-aware super-resolution refiner [1]. The model was trained using a data engine that curated more than 100 million training clips from 50 million raw videos [1]. A five-stage training pipeline includes flow matching pre-training, personality fine-tuning, two-phase distillation yielding greater than 10x acceleration, and RLHF alignment, deployed across thousands of GPUs [1]. Avatar V generates 1080p videos of unlimited duration and achieves state-of-the-art identity preservation, lip synchronization, and generation quality on a cross-scene benchmark, outperforming systems including Seedance 2.0, Kling O3 Pro, Veo 3.1, and OmniHuman 1.5 in both automated metrics and human evaluation [1]. The paper appears on arXiv, a preprint repository that has integrated with platforms such as Hugging Face Spaces to make machine learning research more accessible through interactive demos [8][9]. Hugging Face Spaces allows users to share and explore open-source machine learning demos without writing code, and since November 2022 those demos have been linked directly from arXiv abstract pages via a dedicated Demos tab [8][10]. The collaboration was designed to increase reproducibility and let a wider audience identify biases and other issues in computational models [8]. The release of Avatar V adds to a period of rapid advancement in large-scale AI models. In early 2025, the Chinese firm DeepSeek drew industry attention when it released its DeepSeek-R1 model, which provided responses comparable to OpenAI’s GPT-4 and o1 while reportedly costing far less to train [11]. DeepSeek claimed its V3 model was trained for US$6 million, roughly one-tenth the computing power used by Meta’s comparable Llama 3.1 model [11]. That event sent what observers called “shock waves” through the industry and contributed to a single-day US$600 billion loss in Nvidia’s market value [11].