Capybara is a unified visual creation model that combines advanced diffusion models and transformer architectures for high-quality visual synthesis and manipulation. The framework supports multiple tasks including Text-to-Video, Text-to-Image, and instruction-based editing across images and videos. xgen-universe released version 0.1 with distributed inference support for efficient multi-GPU processing.

Key features include multi-task support covering T2V, T2I, TV2V, and TI2I operations with precise control over content, motion, and camera movements. The framework leverages distributed inference for efficient multi-GPU processing and recently added ComfyUI support with custom nodes for all task types. FP8 quantization support improves inference speed while maintaining quality.

Installation requires CUDA 12.6 and PyTorch 2.6.0, with models organized in a specific directory structure. The framework offers both single sample and batch processing modes via CSV files. ComfyUI integration provides custom nodes for pipeline loading, generation, video handling, and instruction rewriting using Qwen3-VL-8B-Instruct. Capybara represents a significant advancement in unified visual creation systems, offering developers a comprehensive toolkit for AI-powered visual content generation and editing.