SeedVR2 | Image & Video Upscaler

What is SeedVR2?#

SeedVR2 is a revolutionary one-step video restoration model developed by ByteDance Seed, built on a Diffusion Transformer architecture. Unlike traditional diffusion models that require dozens of inference steps, SeedVR2 achieves single-step generation, delivering over 4x faster video restoration while maintaining or even exceeding the quality of multi-step models.

Think of SeedVR2 as a super-intelligent "video repair specialist" that can instantly transform blurry, low-resolution videos and images into crisp, high-quality results with just one pass—no need for the repeated "polishing" that other tools require.

Note: If you want to use video upscaling capabilities and your input video is longer than 30 seconds, please switch to the SeedVR V2.5 workflow and run it on a 2XL or higher machine. For reference, processing a 1-minute video takes under 30 minutes on a 2XL machine, and approximately 20 minutes on a 3XL machine.

Benefits of ComfyUI SeedVR2#

• Lightning-Fast Processing: SeedVR2 achieves one-step restoration, eliminating the need for time-consuming multi-step inference
• Exceptional Enhancement Quality: SeedVR2 delivers stunning results for both image and video restoration tasks
• Dual-Mode Capability: SeedVR2 handles both image and video restoration within a single unified workflow
• Adaptive Resolution Support: SeedVR2 processes arbitrary input sizes without resolution constraints
• Superior Detail Recovery: SeedVR2 excels at recovering fine details like text, facial features, and architectural elements
• Memory Efficient: The SeedVR2 3B model provides excellent quality-to-performance ratio

How to Use ComfyUI SeedVR2 Workflow#

Understanding the SeedVR2 Dual Workflow Structure#

The SeedVR2 workflow consists of two distinct processing paths:

1. Upper Path - SeedVR2 Image Enhancement Workflow
2. Lower Path - SeedVR2 Video Enhancement Workflow

Both SeedVR2 workflows share the same core SeedVR2 technology and can deliver impressive restoration results across different media types.

SeedVR2 Image Enhancement Workflow (Upper Path)#

SeedVR2 Setup Process:

1. Load Image Node: Upload your blurry or low-resolution image for SeedVR2 processing
2. SeedVR2 Video Upscaler Node: Configure your SeedVR2 enhancement settings
   • new_width: Set target width
   • cfg_scale: Use 1.01 for subtle SeedVR2 enhancement or higher values for more dramatic changes
   • batch_size: Keep at 1 for single image SeedVR2 processing
3. LayerUtility Nodes: The workflow includes comparison tools to evaluate SeedVR2 before/after results
4. Save Image: Export your SeedVR2 enhanced result

SeedVR2 Video Enhancement Workflow (Lower Path)#

SeedVR2 Video Setup Process:

1. Load Video Node: Upload your video file for SeedVR2 processing
   • Set appropriate force_rate, custom_width, and custom_height for SeedVR2
   • Configure frame_load_cap to limit SeedVR2 processing for testing
   • Use select_every_nth to skip frames if needed for faster SeedVR2 processing
2. SeedVR2 Video Upscaler: Use identical SeedVR2 settings as image workflow
3. Video Combine Node: Compile SeedVR2 processed frames back into video format
   • frame_rate: Match your input video's framerate for SeedVR2 output
   • format: Choose appropriate output format (mp4 recommended for SeedVR2)
   • crf: Control compression quality (19 is a good balance for SeedVR2)

SeedVR2 Parameter Optimization Guide#

Essential SeedVR2 Settings for Best Results:

SeedVR2 Model Selection:

• seedvr2_ema_3b_fp16.safetensors: Best overall SeedVR2 choice (requires 18GB+ VRAM)
• seedvr2_ema_3b_fp8_e4m3fn.safetensors: Lower memory SeedVR2 alternative with slightly reduced quality
• seedvr2_ema_7b_fp16.safetensors: Larger SeedVR2 model for potentially better quality (higher VRAM requirements)
• seedvr2_ema_7b_fp8_e4m3fn.safetensors: SeedVR2 7B model with reduced precision
• Currently avoid SeedVR2 7B models due to reported vertical stripe artifacts

SeedVR2 Quality Controls:

• cfg_scale: 1.01-1.5 for subtle SeedVR2 enhancement, 2.0-3.0 for aggressive SeedVR2 restoration
• seed: Experiment with different values—each produces unique SeedVR2 detail reconstruction
• preserve_vram: Always enable unless you have abundant GPU memory for SeedVR2

More Information about SeedVR2#

For additional SeedVR2 technical details and research background:

• Original SeedVR2 research by ByteDance Seed Team
• ComfyUI SeedVR2 implementation by NumZ
• SeedVR2 project homepage with demos and comparisons: SeedVR2 Website

Acknowledgements#

This workflow is powered by SeedVR2, developed by the ByteDance Seed research team. The ComfyUI SeedVR2 integration is provided by NumZ, enabling seamless one-step video and image restoration within the ComfyUI ecosystem. Special recognition goes to the original authors for their groundbreaking work in making high-quality SeedVR2 diffusion-based restoration both fast and accessible.