AptAvatar:Fast and Vivid Long-Form Audio-Driven Video Generation for Production-Ready Avatars

Hengyuan Zhang^†Jingna Sun^†Meiguang Jin^✉Junfeng Ma

Taobao & Tmall Group of Alibaba

† Equal contribution✉ Corresponding author

A powerful fine-tuned multi-step generator distilled to two inference steps—while retaining high fidelity, expressive motion, and long-horizon identity consistency.

Paper arXiv ↗Code GitHub ↗Watch demos

14Bfine-tuned generator

720phigh-resolution video

2NFEs at inference

60×inference speedup

166-SECOND OUTPUTExtended long-video demo

2-NFE generation

Identity-consistent

VIDEO DEMOS

Short and long videos

Cartoon short videos, real-person short videos, and minute-level long videos.

SHORT VIDEO · CARTOON

Cartoon generation

10 short-video demos

Wood Puppet15 seconds · stylized avatar

Cel-Shaded Anime15 seconds · stylized avatar

Dollhouse15 seconds · stylized avatar

Clay Stop-Motion15 seconds · stylized avatar

3D Ink Wash15 seconds · stylized avatar

Pastel Diorama15 seconds · stylized avatar

Ink-Wash Toon15 seconds · stylized avatar

Porcelain Figurine15 seconds · stylized avatar

Art Nouveau15 seconds · stylized avatar

Cyberpunk Anime15 seconds · stylized avatar

SHORT VIDEO · REAL PERSON

Expressive real-person generation

25 short-video demos

Real-person demo 0115 seconds · 704×1280

Real-person demo 0215 seconds · 704×1280

Real-person demo 0315 seconds · 704×1280

Real-person demo 0415 seconds · 704×1280

Real-person demo 0515 seconds · 704×1280

Real-person demo 0615 seconds · 704×1280

Real-person demo 0715 seconds · 704×1280

Real-person demo 0815 seconds · 704×1280

Real-person demo 0915 seconds · 704×1280

Real-person demo 1015 seconds · 704×1280

Real-person demo 1115 seconds · 704×1280

Real-person demo 1215 seconds · 704×1280

Real-person demo 1315 seconds · 704×1280

Real-person demo 1415 seconds · 704×1280

Real-person demo 1515 seconds · 704×1280

Real-person demo 1615 seconds · 704×1280

Real-person demo 1715 seconds · 704×1280

Real-person demo 1815 seconds · 704×1280

Real-person demo 1915 seconds · 704×1280

Real-person demo 2015 seconds · 704×1280

Real-person demo 2115 seconds · 704×1280

Real-person demo 2215 seconds · 704×1280

Real-person demo 2315 seconds · 704×1280

Real-person demo 2415 seconds · 704×1280

Real-person demo 2515 seconds · 704×1280

LONG VIDEO

Long-video generation

4 minute-level demos

Long-form demo 0148.5 seconds · 704×1280

Long-form demo 0248.4 seconds · 704×1280

Long-form demo 0340.6 seconds · 704×1280

Long-form demo 0448.4 seconds · 704×1280

MOTIVATION

Production-ready generation requires both speed and long-video quality.

Fast inference must be balanced with motion quality and identity consistency over long videos.

AptAvatar teaser showing reference, audio, and text inputs; vivid short-video motion; and identity-consistent long-video generation — AptAvatar accepts a reference image, speech audio, and fine-grained text prompts, then produces expressive short-form motion and stable long-form video with only two inference steps.

Extreme acceleration, no capacity cut

Existing speedups often causalize attention, shrink the model, or shorten its temporal window. AptAvatar keeps the 14B bidirectional backbone and compresses its denoising trajectory instead.

Long-form generation without drift

Chunk-wise generation must condition on its own imperfect history. Small errors accumulate into identity drift, degraded hands, and reduced motion expressiveness unless training reproduces that deployment reality.

METHOD

Anchor the endpoint. Replay the history.

Two complementary designs make aggressive two-step distillation and stable autoregressive rollout work together.

Endpoint-Anchored Distribution Distillation

A frozen four-step bridge generator defines an attainable trajectory endpoint. A dedicated Anchor Score Estimator turns that fixed endpoint into stable guidance for the evolving two-step student.

EADD

Self-Generated History Replay

Cached outputs from earlier generator checkpoints become history conditions during training, approximating inference-time self-conditioning without costly recurrent online rollouts.

SGHR

QUANTITATIVE