seedvr2/upscale/video

Introduction of SEEDVR2

SeedVR2 is a breakthrough one-step video restoration tool created by ByteDance Seed, featuring a Diffusion Transformer design. It delivers lightning-fast performance, restoring clarity in a single pass while preserving fine detail. With SeedVR2 video and Video Upscale models, you can enhance blurry content into high-quality media efficiently and at scale.

SEEDVR2 empowers creators, editors, and professionals who need instant, high-quality restoration of videos. It transforms noisy or low-resolution media into sharp outputs with remarkable efficiency. Whether you work with old footage, compressed files, or everyday videos, SEEDVR2 enhances details and sharpness while generating output at a professional quality standard.

Key Models for SEEDVR2

seedvr2_ema_3b_fp16

This model is the best overall choice for SeedVR2, recommended when abundant GPU memory is available. It ensures the highest quality restoration and excels in recovering fine details such as text, faces, and architecture. With one-step generation, it maintains speed while achieving excellence in visual quality.

seedvr2_ema_3b_fp8_e4m3fn

This configuration offers similar capability to the 3B FP16 model while reducing memory requirements. It is particularly useful if you are working on a system with lower VRAM, trading slight quality reduction for much better accessibility in memory-limited environments.

seedvr2_ema_7b_fp16

Designed as a larger SeedVR2 model, the 7B FP16 option can potentially deliver higher quality restoration results compared to the 3B models. However, it requires significantly higher VRAM and has known issues with vertical stripe artifacts, so it is less recommended for production use.

seedvr2_ema_7b_fp8_e4m3fn

This compressed precision version of the 7B model provides reduced memory usage while attempting to preserve detail quality. While it supports advanced experimentation, vertical stripe artifacts remain a disadvantage, making it best suited for testing scenarios rather than critical restoration tasks.

How to Use SEEDVR2

Inputs Required

You must provide an video input to begin processing. This file is the source that SEEDVR2 will upscale and restore. Without this, SEEDVR2 cannot operate, as the tool requires content to enhance from its original state.

Optional Inputs and Controls

You can adjust the model and CFG parameter to influence the unique detail reconstruction effects during upscaling. Additionally, you may define New Width to control the size of final output. These adjustments let you fine-tune results for consistency and playback quality.

Outputs

Once processed, SEEDVR2 produces enhanced videos that are sharper and more detailed than the input.

Best Practices

For CFG, use 1.01 for subtle SeedVR2 enhancement or higher values for more dramatic changes. Begin with default settings and adjust only if needed to balance quality and efficiency for your project.