我们提出了 HunyuanVideo 1.5，一个轻量级但功能强大的开源视频生成模型，仅用 83 亿参数就实现了最先进的视觉质量和运动连贯性，能够在消费级 GPU 上进行高效推理。这一成就建立在几个关键组件之上，包括精心设计的数据整理、采用选择性滑动窗口注意力（SSTA）的高级 DiT 架构、通过字形感知文本编码增强的双语理解、渐进式预训练和后训练，以及高效的视频超分辨率网络。利用这些设计，我们开发了一个统一的框架，能够跨多种时长和分辨率进行高质量的文本到视频和图像到视频生成。大量实验表明，这个紧凑而高效的模型在开源视频生成模型中树立了新的标杆。通过发布代码和模型权重，我们为社区提供了一个高性能基础，降低了视频创作和研究的门槛，使先进的视频生成技术惠及更广泛的受众。所有开源资源均可公开获取：https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5。
We present HunyuanVideo 1.5, a lightweight yet powerful open-source video generation model that achieves state-of-the-art visual quality and motion coherence with only 8.3 billion parameters, enabling efficient inference on consumer-grade GPUs. This achievement is built upon several key components, including meticulous data curation, an advanced DiT architecture featuring selective and sliding tile attention (SSTA), enhanced bilingual understanding through glyph-aware text encoding, progressive pre-training and post-training, and an efficient video super-resolution network. Leveraging these designs, we developed a unified framework capable of high-quality text-to-video and image-to-video generation across multiple durations and resolutions. Extensive experiments demonstrate that this compact and proficient model establishes a new state-of-the-art among open-source video generation models. By releasing the code and model weights, we provide the community with a high-performance foundation that lowers the barrier to video creation and research, making advanced video generation accessible to a broader audience. All open-source assets are publicly available at https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5.
核心贡献 · Key contributions
提出轻量级 8.3B 参数 DiT 与 SSTA，在消费级 GPU 上高效生成视频。 Proposes a lightweight 8.3B-parameter DiT with SSTA for efficient video generation on consumer GPUs.
引入两阶段流程：统一 DiT 用于文生视频/图生视频，以及视频超分网络提升至 1080p。 Introduces a two-stage pipeline: unified DiT for T2V/I2V and a video super-resolution network for 1080p upscaling.
开发选择性滑动块注意力（SSTA）剪枝冗余 token，实现 1.87 倍加速。 Develops Selective and Sliding Tile Attention (SSTA) to prune redundant tokens, achieving 1.87x speedup.
通过 Qwen2.5-VL 和 Glyph-ByT5 文本编码器增强双语理解，实现精准文本渲染。 Enhances bilingual understanding via Qwen2.5-VL and Glyph-ByT5 text encoders for accurate text rendering.
采用渐进式预训练和后训练（CT、SFT、RLHF）提升运动连贯性与对齐。 Employs progressive pre-training and post-training (CT, SFT, RLHF) to improve motion coherence and alignment.
以紧凑的 8.3B 参数实现开源视频生成最先进水平，降低研究门槛。 Achieves state-of-the-art open-source video generation with compact 8.3B parameters, lowering barriers for research.
局限 · Limitations
模型性能仅在精选提示上评估，在分布外输入上可能下降。 Model performance is evaluated only on curated prompts and may degrade on out-of-distribution inputs.
推理速度测量使用未优化研究代码；实际部署可能需要进一步工程优化。 Inference speed measurements use unoptimized research code; real-world deployment may require further engineering.
超分网络在 100 万片段上训练；在未见分辨率或时长上的性能未经测试。 The super-resolution network is trained on 1M clips; performance on unseen resolutions or durations is untested.
稀疏注意力（SSTA）仅在蒸馏阶段应用；无它时完整训练可能产生不同结果。 Sparse attention (SSTA) is applied only during distillation; full training without it may yield different results.
模型开源但依赖专有数据整理和标注流程，限制可复现性。 The model is open-source but relies on proprietary data curation and captioning pipelines, limiting reproducibility.
论文章节 · Sections（共 16）
摘要Abstract
1 引言1 Introduction
2.1 数据获取与过滤2.1 Data Acquisition and Filtering
2.2 字幕生成2.2 Captioning
3 模型设计3 Model Design
3.1 统一扩散变换器3.1 Unified Diffusion Transformer
3.2 视频超分辨率3.2 Video Super-Resolution
4.1 预训练4.1 Pre-training
4.2 后训练4.2 Post-training
4.3 视频超分辨率4.3 Video Super-Resolution.
5 模型性能5 Model Performance
5.1 评分5.1 Rating
5.2 GSB5.2 GSB
6 推理速度与 GPU 内存需求6 Inference Speed and GPU Memory Requirements