smthemex/ComfyUI_FlashVSR

What it solves

This project provides a ComfyUI integration for FlashVSR and FlashVSR+, which are advanced diffusion-based frameworks designed for high-quality video super-resolution. It enables users to upscale videos efficiently, including support for 4K resolution processing on consumer hardware like the RTX 4090.

How it works

The project acts as a custom node wrapper for ComfyUI, allowing users to load FlashVSR models and run them within their existing ComfyUI workflows. It supports multiple attention mechanisms, including Block-Sparse-Attention and Triton-based FlashVSR attention, to balance speed and memory usage. Users can configure parameters like input frame counts and VAE inference modes to optimize for either speed or output quality.

Who it’s for

This tool is intended for video editors, AI artists, and researchers using ComfyUI who need to perform high-quality video upscaling using state-of-the-art diffusion models.

Highlights

• Enables real-time, diffusion-based video super-resolution within ComfyUI.
• Supports FlashVSR+ for efficient, high-quality 4K video upscaling.
• Offers flexible attention configurations to manage VRAM usage and prevent out-of-memory (OOM) errors.
• Provides options for different VAE modes to trade off between processing speed and visual fidelity.