XGen - Video Generation Platform

2025Case Study

Scalable text-to-video and image-to-video pipeline optimized for realism and reproducibility.

The Challenge

  • Existing pipelines struggle with realism and temporal consistency
  • Scaling content production requires reproducible workflows

The Solution

  • ComfyUI-based DAG workflows running on GPU providers

  • Open-source models with LoRAs, VAEs, ControlNet conditioning

  • Storage and delivery optimized for scale

Impact

  • Demonstrated commercial viability of AI video generation at scale

Technical

Implementation Details

Wan 2.1 models, ComfyUI DAG export for reproducibility, GPU providers (Modal, Runpod), Backblaze B2 storage, Bunny CDN delivery, observability for inference time and errors.