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

Links are placeholders for now and will be updated as project pages, model cards, community channels, and reports become available.

✨ Key Features

• Endpoint-Anchored Distribution Distillation: It augments vanilla distribution matching with a dedicated Anchor Score Estimator trained on the trajectory-endpoint distribution defined from a frozen pretrained 4-step bridge generator. This provides an attainable endpoint-level anchor for the evolving two-step student.
• Self-Generated History Replay: It reuses cached outputs from earlier generator checkpoints as history conditions during chunk-wise training. This approximates inference-time self-conditioning, mitigating quality degradation from accumulated history errors

🔥 News

• 2026.06.29 - Released the initial AptAvatar repository with open-source inference code.

📋 Todo

• Technical report
• Project page
• Inference code
• Checkpoint release
• Additional demos

🧩 Pipeline

🎬 Examples

🚀 Quick Start

🛠️ Installation

Create and activate the environment:

conda create -n aptavatar python=3.10
conda activate aptavatar

Install PyTorch:

pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128

Install project dependencies:

pip install -r requirements.txt

Install FlashAttention:

pip install ninja
pip install flash_attn==2.8.0.post2 --no-build-isolation

If flash_attn_3 is installed in your environment, please uninstall it, as it conflicts with the FlashAttention-3 kernels bundled in xformers:

pip uninstall flash_attn_3

Install FFmpeg:

# Ubuntu / Debian
apt-get install ffmpeg

# CentOS / RHEL
yum install ffmpeg ffmpeg-devel

or install through conda:

conda install -c conda-forge ffmpeg==7

📦 Model Preparation

The current repository provides inference code only. After checkpoints are released, place them under the following paths:

Component	Expected Path	Status
AptAvatar checkpoint	models/AptAvatar	AptAvatar
Audio encoder	models/chinese-wav2vec2-base	chinese-wav2vec2-base

# Placeholder commands. Replace the repository names after release.
pip install "huggingface_hub[cli]"
huggingface-cli download TaoLiveAIGC/AptAvatar --local-dir ./models/AptAvatar
huggingface-cli download TencentGameMate/chinese-wav2vec2-base --local-dir ./models/chinese-wav2vec2-base

🎥 Inference

Single-GPU inference:

bash inference_script_single_gpu.sh

Multi-GPU inference:

bash inference_script_multi_gpu.sh

The default scripts use:

examples/woman.png
examples/woman.wav
examples/woman.txt

You can edit the shell scripts or call generate_video.py directly with your own image, audio, and prompt.

📝 Structured Prompt

Besides free-form prompts, AptAvatar supports structured prompts that describe the motion step by step with explicit frame ranges. See examples/woman_template.txt for full examples. Each prompt follows the format:

步骤1：*帧 <start>~<end>*<motion description>步骤2：*帧 <start>~<end>*<motion description>步骤3：*帧 <start>~<end>*<motion description>

• Each step starts with 步骤x： followed by a frame range marker *帧 xx~xx* and a motion description.
• We recommend using 1~3 steps per prompt.

Example:

步骤1：*帧 0~20*女主播面向镜头站立，右手先抬起至胸前高度并张开手掌，掌心朝前，身体保持稳定，视线面向镜头。步骤2：*帧 25~45*她的右手由张开状态逐渐收拢成松拳，手腕和手指动作清晰，头部和身体保持正向稳定。步骤3：*帧 50~70*女主播左手随后抬起至与右肩相近的高度，双手同时张开并向身体两侧展开，形成对称的展示姿态后保持定格。

📊 Evaluation Results

We evaluate AptAvatar on our internal short-form benchmark for audio-driven avatar generation. Results will be reported at 720p.

Metrics

Category	Metrics	Direction
Audio-Visual Alignment	Sync-C, Sync-D	Sync-C higher is better; Sync-D lower is better
Image Quality	FID	Lower is better
Aesthetic Analysis	ASE, Q-Align	Higher is better
VBench	Subject Consistency (Subject-C), Background Consistency (BG-C), Motion Smoothness (Motion-S), and Temporal
Flickering (Temporal-F)	higher is better

Compared Methods

Group	Method	NFE
Many-NFE	InfiniteTalk	120
Many-NFE	Wan-S2V	80
Few-NFE	LongCat-Video-Avatar 1.5	8
Few-NFE	LiveAvatar	4
Few-NFE	SoulX-FlashTalk	4
Few-NFE	AptAvatar	2

720p Results on our internal short-form benchmark

Method	Group	NFE	Sync-C ↑	Sync-D ↓	FID ↓	ASE ↑	Subject-C ↑	BG-C ↑	Motion Smoothness ↑	Temporal Flickering ↑
InfiniteTalk	Many-NFE	120	7.172	8.010	24.808	3.146	96.889	95.568	99.512	99.318
Wan-S2V	Many-NFE	80	5.626	9.376	43.542	3.119	97.151	95.663	99.385	99.216
LongCat-Video-Avatar 1.5	Few-NFE	8	4.140	11.433	36.582	3.191	96.338	94.533	99.550	99.332
LiveAvatar	Few-NFE	4	5.986	8.586	37.868	3.184	97.804	95.718	99.409	99.226
SoulX-FlashTalk	Few-NFE	4	7.743	7.912	28.039	3.172	98.580	96.530	99.550	99.434
AptAvatar	Few-NFE	2	7.865	7.739	24.759	3.307	98.589	96.844	99.561	99.456

📁 Repository Layout

AptAvatar/
├── README.md
├── LICENSE
├── CHANGELOG.md
├── requirements.txt
├── config.json
├── generate_video.py
├── gradio_app.py
├── inference_script_single_gpu.sh
├── inference_script_multi_gpu.sh
├── examples/
│   ├── woman.png
│   ├── woman.wav
│   └── woman.txt
├── assets/
│   └── *.mp4
├── sample_results/
│   └── *.mp4
├── models/                                  # created after downloading checkpoints
│   ├── AptAvatar/
│   │   ├── config.json
│   │   ├── configuration.json
│   │   ├── diffusion_pytorch_model-0000*-of-00008.safetensors
│   │   ├── diffusion_pytorch_model.safetensors.index.json
│   │   ├── Wan2.1_VAE.pth
│   │   ├── models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth
│   │   ├── models_t5_umt5-xxl-enc-bf16.pth
│   │   ├── google/
│   │   │   └── umt5-xxl/                    # T5 tokenizer files
│   │   └── xlm-roberta-large/               # CLIP tokenizer files
│   └── chinese-wav2vec2-base/
│       ├── config.json
│       ├── preprocessor_config.json
│       └── pytorch_model.bin
└── AptAvatar/
    ├── inference.py
    ├── configs/
    ├── infinite_talk/
    ├── src/
    └── wan/

⚖️ License Agreement

This repository is released under the Apache License 2.0. See LICENSE for the full license text.

The current release contains inference code. Model weights, datasets, benchmark assets, demo media, and third-party dependencies may be released under separate terms when they become available. Users are responsible for checking and complying with the licenses of all external models, datasets, audio encoders, and media assets used with this repository.

⚠️ Usage Considerations

AptAvatar is intended for research and responsible application development in audio-driven avatar video generation. Users should carefully evaluate outputs before deployment, especially in identity-sensitive, public-facing, or commercial scenarios.

• Obtain appropriate rights and consent for all input images, audio, prompts, and generated likenesses.
• Do not use the system to impersonate real people, mislead audiences, or generate deceptive content.
• Clearly disclose synthetic or AI-generated avatar content when required by law, platform policy, or product context.
• Test robustness across languages, accents, speaking styles, lighting conditions, identities, and long-form generation settings before production use.
• Follow applicable laws and regulations related to privacy, biometric data, publicity rights, copyright, and content safety.

📚 Citation

Citation information will be added after the technical report is released.

@misc{taoavatar2026,
  title        = {AptAvatar: Fast and Vivid Long-Form Audio-Driven Video Generation for Production-Ready Avatars},
  author       = {TBD},
  year         = {2026},
  archivePrefix = {arXiv},
  eprint       = {TBD},
  primaryClass = {cs.CV},
  url          = {TBD}
}

🙏 Acknowledgements

AptAvatar builds on the progress of the open-source video generation and audio-driven avatar communities. We sincerely thank the authors and contributors of the following projects:

• InfiniteTalk and Wan: foundational video generation and audio-conditioned avatar backbones that AptAvatar builds upon.
• Self-Forcing: an important reference codebase for long-horizon self-conditioned video generation.
• DMD and Self-Forcing++: key references for distribution matching distillation and self-forcing-style training.

If you find AptAvatar useful, please also consider starring the original repositories of these foundational methods.