> For the complete documentation index, see [llms.txt](https://nadai.gitbook.io/nadsai-whitepaper/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://nadai.gitbook.io/nadsai-whitepaper/nad101-technical-architecture.md). # Nad101 Technical Architecture How a video is generated by Nad101

#### **1. Text Input and Preprocessing** User Input: Users submit a text prompt. LLM Text Extending: Extends the text descriptions to add context and detail. LLM Re-captioning: Generates detailed captions for visual content. #### **2. Video Encoding and Diffusion** Visual Encoder (VAE Encoder E): Encodes the training video into visual patches E(x). Diffusion Process: Applies diffusion to the visual patches, resulting in diffused visual patches. #### **3. Latent Representation and Conditioning** ViT Patch + Position: Combines diffused visual patches with positional information to form latent representations z. Conditional Latent zc: Enhances Latent z with contextual information from images, videos, and text descriptions. #### **4. Transformer Encoding and Decoding** Transformer Encoder (DiT): Processes the conditional latents through a series of transformer encoder layers. Linear Decoder and Reshape: Decodes and reshapes the transformed latents into visual patches. #### **5. Video Reconstruction and Output** VAE Decoder: Decodes the visual patches into the final predicted video. Rendering and Optimization: Renders the video in high resolution and optimizes it for various formats. *This detailed process ensures Nad101 generates high-quality, contextually accurate videos efficiently using advanced AI techniques and decentralized data sources.*