Workflow for video upscaling and interpolation

Designed for short 5-6s videos. Use at your own risk for longer videos (more frames = more RAM required).

v1 Example was generated from 480x832 12fps video using Seed VR2 at 3x scale and interploated to 60fps with 5x multiplier.

v2 update - Color Match / Image-based Color Grading

• Added Color Match as an optional color correction feature, mainly to fix color deviation from GGUF models and ESRGAN upscalers.
  

• Color Match can also be used as a color grading tool, for both ESRGAN-based upscale and SeedVR2.
  

• Scale is now always set correctly, regardless of ESRGAN model being used.

To use Color Match, first, upload an original video and an image used to generate it, or an image for color grading.

Then, in ESRGAN upscaling enable both color_correct and color_match. Color Blend settings will be ignored when using Color Match. In SeedVR2 upscaler there's only one toggle to enable Color Match.

(Color Match only) If you're experiencing lag during upscaling, especially at 3x+ scale, disable multithread to reduce CPU load.

Getting started

• Install missing nodes using ConmyUI Node Manager.

• Seed VR2 should download necessary models automatically after custom nodes are installed.

• For best results do not use codec compression when saving your low res generated videos.

Upscale with ESRGAN

(v1 only) This part of a workflow assumes ESRGAN model generates images at 4x scale (i.e. - Remacri, 4x-UltraSharp) by default and calculates final resolution accordingly.

Pros

Fast, less VRAM and RAM hungry. I upscale to 1440 on a 12GB VRAM card with 32GB system RAM without a hitch.

Cons

Medium quailty, depends on an ESRGAN used.

Upscale with SeedVR2

Pros

Considerably higher quality for realistic videos.

Cons

Much slower, very VRAM and RAM hugnry at a resolution higher than 1080p.

RIFE Interpolation

Generates additional frames. Use rife49 or rife47 models for best results (49 is set as default).

Save final frames

Saves last N frames if you plan on reusing them for the next I2V generation or First to Last frame workflow.