Skip to main content

​
ComfyUI SeedVR2 Introduction

SeedVR2 (arXiv:2506.05301, from ByteDance Seed) is a one-step diffusion-based video restoration model. It performs adversarial training against real data to restore and upscale images and videos with high fidelity. SeedVR2 is natively supported in ComfyUI (PR #14424). Key capabilities:
  • Conservative upscaling: preserves original structure and detail while enhancing clarity
  • One-step inference: single forward pass for both image and video upscaling
  • Multiple model sizes: 3B and 7B variants with FP16, FP8, INT8, NVFP4, and MXFP8 quantizations
  • Both image and video: supports single-image upscaling and video resolution enhancement
SeedVR2 comes in two sizes. The 3B model is a smaller variant with lower VRAM requirements, suitable for users with limited hardware. The 7B model is a larger variant that delivers higher quality results at the cost of increased VRAM and processing time. Both are available under the Apache 2.0 license.
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you can’t find them in the template, your ComfyUI may be outdated.If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

​
Model Installation

Download the SeedVR2 checkpoint(s) from the Comfy-Org SeedVR2 repository and save them to the corresponding ComfyUI folder: Place the files in the following directories:

​
Example Workflows

​
SeedVR2 3B Int8: Upscale Image

Upscale images using SeedVR2 3B Int8, a one-step diffusion-based video restoration model that produces high-quality results with improved temporal consistency. SeedVR2 3B Int8 upscale image workflow preview

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search “SeedVR2 3B Int8: Upscale Image” in Template Library
Input materials Upload this file to the matching LoadImage node:

watch_macro_shot.png

LoadImage node 1 · watch_macro_shot.png
Example output
Input imageSeedVR2 3B Int8 upscale example output

​
1.1 Steps to Run

  1. Place your image in the ComfyUI/input/ folder and select it in the Load Image node
  2. Select the seedvr2_3b_int8_convrot.safetensors checkpoint in the SeedVR2 model loader
  3. Click Queue or use Ctrl(cmd) + Enter to run

​
SeedVR2 7B Int8: Upscale Image

Upscale images using SeedVR2 7B Int8, a one-step diffusion model that enhances resolution through adversarial training and adaptive window attention. SeedVR2 7B Int8 upscale image workflow preview

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search “SeedVR2 7B Int8: Upscale Image” in Template Library
Input materials Upload this file to the matching LoadImage node:

indoor_portrait.png

LoadImage node 1 · indoor_portrait.png
Example output
Input imageSeedVR2 7B Int8 upscale example output

​
2.1 Steps to Run

  1. Place your image in the ComfyUI/input/ folder and select it in the Load Image node
  2. Select the seedvr2_7b_int8_convrot.safetensors checkpoint in the SeedVR2 model loader
  3. Click Queue or use Ctrl(cmd) + Enter to run

​
SeedVR2 3B Int8: Upscale Video

Upscale and restore video footage using SeedVR2 3B Int8, a one-step diffusion model that enhances resolution while maintaining temporal consistency across frames. SeedVR2 3B Int8 upscale video workflow preview

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search “SeedVR2 3B Int8: Upscale Video” in Template Library
Input materials Upload this file to the matching LoadVideo node:

grainy_perfume_shot_crf32.mp4

LoadVideo node 73 · grainy_perfume_shot_crf32.mp4
Example output

​
3.1 Steps to Run

  1. Place your video in the ComfyUI/input/ folder and select it in the Load Video node
  2. Select the seedvr2_3b_int8_convrot.safetensors checkpoint in the SeedVR2 model loader
  3. Click Queue or use Ctrl(cmd) + Enter to run

​
Performance

Higher target resolutions require more processing time. The INT8 variant provides efficient inference with reduced VRAM usage compared to FP16.

​
Troubleshooting

​
The upscaled image is fully black

Attention backends can silently return black or corrupted output on quantized checkpoints. If the SeedVR2 INT8 workflows produce an all-black image, try these in order:
  1. Run without FlashAttention: if you enabled it through --use-flash-attention or a custom node pack, start ComfyUI with the default attention backend and re-run. Disabling FlashAttention has resolved all-black output on the 3B and 7B INT8 workflows on an RTX 3090.
  2. Switch the checkpoint: use the FP16 or FP8 file of the same size. If those render correctly, the INT8 quantization path is where it breaks.
  3. Verify the files: the diffusion model and seedvr2_ema_vae_fp16.safetensors must both be fully downloaded and placed as shown in Model Installation. A truncated download fails the same way.
  4. Disable the live preview: in some cases only the preview is black while the saved file is correct, so turning the live preview off removes the ambiguity.
If the result is still black, post on the tutorial discussion with your GPU, ComfyUI version, install method (Desktop, portable, or manual) and the console output, plus whether the 3B and 7B workflows fail identically.

​
Community Resources