Avatar V: Scaling Video-Reference Avatar Video Generation

2026-06-11

AI Summary

This work addresses the limitations of existing avatar generation methods that rely on static images for identity conditioning, which struggle to capture dynamic behavioral traits—such as speech rhythm, micro-expressions, and gestures—and suffer from perceptual quality degradation in facial regions due to pixel-level losses. To overcome these challenges, we propose the first video-reference-based avatar synthesis framework that jointly learns static identity and dynamic behavior by modeling full video token sequences, enabling generation of arbitrarily long 1080p videos. Key innovations include a sparse reference attention mechanism for linear-complexity long-video modeling, a motion representation stream with closed-loop speaking style transfer, and an identity-aware super-resolution refinement module. Integrated with a five-stage training pipeline and a large-scale data engine, our method significantly outperforms state-of-the-art systems—including Seedance 2.0 and Kling O3 Pro—on cross-scenario benchmarks, achieving superior performance in identity fidelity, lip-sync accuracy, and overall generation quality.

Abstract

Generating avatar videos that are not merely visually similar to a target individual but behaviorally recognizable, faithfully reproducing their talking rhythm, gestural tendencies, and expression dynamics, remains an open challenge. Existing methods predominantly condition on single static images, which provide insufficient identity information and cannot capture dynamic motion traits, while standard pixel-level objectives underserve the perceptually critical facial regions that determine avatar fidelity. We present Avatar V, a production-scale framework that addresses these limitations through video-reference-conditioned identity modeling. Rather than compressing identity into fixed-size embeddings, the model conditions directly on the full token sequence of a reference video, learning to reproduce both static identity attributes (facial geometry, skin texture) and dynamic behavioral patterns (talking rhythm, micro-expressions) through attention over the reference context. We introduce Sparse Reference Attention, an asymmetric mechanism achieving linear-complexity conditioning on arbitrarily long references; a motion representation stream enabling closed-loop talking style transfer; and an identity-aware super-resolution refiner inheriting the full reference conditioning. These are supported by a data engine curating 100M+ training clips from 50M raw videos, and a five-stage training pipeline with flow matching pre-training, personality fine-tuning, two-phase distillation (>10x acceleration), and RLHF alignment, deployed across thousands of GPUs. Avatar V generates 1080p videos of unlimited duration, achieving state-of-the-art identity preservation, lip synchronization, and generation quality on our cross-scene benchmark, consistently outperforming leading systems including Seedance 2.0, Kling O3 Pro, Veo 3.1, and OmniHuman 1.5 in both automated metrics and human evaluation.

Problem

• avatar video generation
• behavioral recognizability
• dynamic motion traits
• identity preservation
• video-reference conditioning

Innovation

• video-reference conditioning
• Sparse Reference Attention
• motion representation stream
• identity-aware super-resolution
• flow matching pre-training