PixelWizard

PixelWizard is a research project for generating videos from text descriptions at unusually high resolutions, specifically 2K (2560x1440) and 4K (3840x2144). Most AI video generation systems produce lower-resolution output because generating high-resolution video is computationally expensive. This project proposes a way to make that process more practical.

The approach works in two stages. First, the system generates a lower-resolution version of the video to establish the overall structure, motion, and timing. Then it generates a high-resolution version, but instead of running the expensive high-resolution process from scratch for every frame, it uses a technique called shortcut step-size conditioning to skip many of the generation steps. The README describes this as decoupling global structure modeling from high-resolution detail generation.

To use PixelWizard, you download two sets of model weights: the base Wan2.2 video generation model (a pre-existing open model the project builds on) and the PixelWizard-specific checkpoints for 2K or 4K generation. You then run a Python script with a text file containing your prompts, and it saves the resulting videos. The hardware requirements are significant: single-GPU inference needs roughly 52 GB of GPU memory for 2K video and about 100 GB for 4K. A multi-GPU mode is available that distributes the memory load across several graphics cards.

This is an early release tied to a research paper posted on arXiv. At the time the README was written, the project page, demo videos, and full paper details were listed as coming soon. The code structure suggests it is intended primarily for researchers and ML engineers rather than general users, given the hardware requirements and the manual setup process involving conda environments, specific PyTorch versions matched to CUDA, and separate checkpoint downloads.

PixelWizard was developed by a team at VisionForge and acknowledges the Wan team for the underlying video generation infrastructure it relies on.