HunyuanVideo operates as a comprehensive AI video generation framework, integrating both text-to-video (T2V) and image-to-video (I2V) functionalities into a unified pipeline. Its core is built upon a systematic framework for large video generation, employing a "Dual-stream to Single-stream" hybrid model design for highly efficient video creation. This innovative architecture processes video and text tokens independently before fusing them, enabling complex interactions between visual and semantic information. The model is trained on a spatial-temporally compressed latent space using a Causal 3D VAE. Text prompts, encoded by a large language model, serve as conditions for generating output latents, which are then decoded into high-quality videos or images. Users can leverage HunyuanVideo's capabilities through various features. The unified generative architecture, based on a Transformer design with Full Attention, handles both image and video generation. An MLLM Text Encoder and 3D VAE are crucial for processing text prompts and compressing pixel-space data. The "Prompt Rewrite" feature offers "Normal" and "Master" modes to refine user intent, enhancing visual quality, composition, lighting, and camera movement. The model boasts strong instruction following, accurately executing bilingual prompts for reliable scene control. It generates natural cinematic camera movements like pans, dollies, tracking shots, and depth shifts. Smooth motion generation and physics compliance ensure fluid movements and reduced visual inconsistencies. Expression fidelity captures nuanced human movements and emotional nuances in real-time. HunyuanVideo supports multi-style generation, allowing users to create videos ranging from realistic scenes to artistic expressions, with seamless transitions between real and virtual styles. It also supports text rendering within generated videos and maintains high image-video consistency. The optimized lightweight architecture, such as HunyuanVideo 1.5 with its 8.3B DiT and 3D causal VAE, ensures efficient, high-quality generation. Accelerated Long-Video Inference, utilizing SSTA (Sliding Tile Attention), reduces redundant attention blocks, significantly boosting inference speed for longer video sequences. A 1080p Video Super-Resolution Enhancement (VSR) module upscales outputs to 1080p, improving clarity for both T2V and I2V results. The platform aims for real-time processing, delivering high-quality video outputs in seconds. HunyuanVideo is an open-source model, generally free for commercial use, though API usage on integrated platforms may incur costs. It supports versatile output specifications, including various resolutions (480p, 580p, 720p, and up to 1080p with VSR), aspect ratios (16:9, 9:16), and video lengths typically ranging from 5 to 10 seconds (85 or 129 frames).