SeedVR2 Long Video Upscaler — HD → 4K

This workflow turns ComfyUI’s native SeedVR2 video upscale into an automatic long-video pipeline.

Instead of loading an entire video into GPU memory or manually processing one short segment at a time, VideoHelperSuite Meta Batch advances through the source automatically until EOF. A custom overlap helper carries boundary frames into the next batch so they are regenerated with future-frame context and written only once.

The result is a workflow designed for arbitrarily long videos, where memory usage depends on the chunk size, not the overall runtime of the source clip.

Highlights

• Arbitrarily long source videos from a VRAM perspective

• Automatic SeedVR2-friendly 4n+1 batch sizing from a simple chunk_seconds control

• 8-frame contextual overlap by default

• No duplicate frames at chunk boundaries

• 2x upscale default: 1080p → 2160p / 4K UHD

• H.264 yuv420p10le, CRF 12 default output

• Original audio restored after processing

• No second video encode during final audio remux

• Optional inner latent splitting for additional VRAM savings

• Optional ProRes output for editing/intermediate use

Default settings

• Chunk duration: 2.0 s

• Outer overlap: 8 frames

• Scale: 2x

• Split latent: off

• Color correction: LAB

• Codec: H.264

• Pixel format: yuv420p10le

• CRF: 12

• Final container: MKV

Requirements

• Current ComfyUI with native SeedVR2 support

• ComfyUI-VideoHelperSuite

• FFmpeg

• SeedVR2 model files:

  • seedvr2_3b_int8_convrot.safetensors

  • seedvr2_ema_vae_fp16.safetensors

Installation

Clone the custom node directly from GitHub into your ComfyUI/custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/biogoly/ComfyUI-SeedVR2-LongVideo.git

Then restart ComfyUI.

After restart, load the included workflow:

ComfyUI-SeedVR2-LongVideo/workflows/SeedVR2_LongVideo_Upscale_v1.4.json

Repo here:

https://github.com/biogoly/ComfyUI-SeedVR2-LongVideo

OOM tuning

If you run out of VRAM:

1. Lower chunk_seconds first

2. If needed, enable split_latent as the secondary memory-control option

This workflow is meant to avoid duration-driven OOM by processing the video in streamed chunks instead of one massive pass.