Alayarenderer Generative World

This is a two-stage video rendering pipeline that first extracts G-buffers (albedo, normal, depth, roughness, metallic) from RGB video using a fine-tuned 7B Cosmos model, then resynthesizes stylized video from those G-buffers plus text prompts using a 1.3B Wan2.1 diffusion model. Think game footage relighting: you can take existing gameplay and transform it to look like it's set in a frozen tundra or cyberpunk rainy night just by changing the prompt. Requires two separate conda environments because of dependency conflicts, which is a bit annoying but manageable. Trained on 4M+ frames from Cyberpunk 2077 and Black Myth: Wukong, so it understands game rendering conventions well.