4K AI Video Upscaler for ComfyUI | LongCat Video Workflow

Introduction to Video Upscaling

AI video generation is remarkable—but raw output often falls short of broadcast quality. Generated videos typically come at 480p (852×480) or 720p (1280×720), with occasional frame stutter and facial artifacts. The solution is comprehensive post-processing: upscaling to higher resolution, restoring facial details, interpolating frames for smoothness, and color correction.

ComfyUI's video upscaler workflow combines cutting-edge AI models to transform rough footage into professional deliverables. In one integrated pipeline, you can take a WAN 2.2-generated video and deliver a crystal-clear 4K masterpiece with frame-interpolated smoothness that looks like it was filmed on $100K+ cinema equipment.

The breakthrough is efficiency. Traditional video upscaling was limited to cloud services or prohibitively expensive hardware. This workflow runs on modest GPUs—even 6GB VRAM systems with batch processing tricks.

Why Professional Upscaling Matters

The Post-Processing Pipeline Impact

Raw AI Video Issues:

• Native resolution: 480p or 720p
• Framerate: 16 FPS (noticeably choppy)
• Facial details: Soft, sometimes artifact-prone
• Color consistency: May drift slightly
• Compression: YouTube re-compresses, losing fine detail

After Upscaling Pipeline:

• Resolution: 1440p or 4K
• Framerate: 60 FPS (buttery smooth)
• Facial details: Sharp, artifact-free with CodeFormer
• Color: Corrected and consistent
• Compression resilience: Survives YouTube re-encoding

Real-World Quality Improvement

Before Upscaling:

• Viewing distance: Close-up (computer monitor)
• Context: Personal projects, social media
• Confidence: "Okay, people can tell it's AI"

After Upscaling:

• Viewing distance: TV, theatrical projection
• Context: Professional portfolios, client work
• Confidence: "Looks professionally shot"

The transformation is night-and-day. Upscaling isn't cosmetic—it's making the difference between AI video as a novelty versus AI video as a professional tool.

System Requirements

Minimum Setup (Works With Lower VRAM)

• GPU: NVIDIA GTX 1080 (8GB) or RTX 3060 (12GB)
• System RAM: 16GB DDR4
• Storage: 50GB free (video cache can be large)
• Batch Size: 5-10 frames per batch
• Processing Time: ~20 minutes per minute of video

Recommended for Comfortable Workflow

• GPU: RTX 4070 (12GB) or RTX 4080 (16GB)
• System RAM: 32GB DDR5
• Storage: 100GB+ free (batch files)
• Batch Size: 20-30 frames per batch
• Processing Time: ~5-8 minutes per minute of video

Premium / Professional Setup

• GPU: RTX 4090 (24GB) or A100 (40GB)
• System RAM: 64GB+
• Storage: NVMe SSD with 200GB+
• Batch Size: 60+ frames at once
• Processing Time: ~2-3 minutes per minute of video

Cloud GPU (Best for One-Off Projects)

• RunPod RTX 4090: $0.44/hour (~2 min per min of video)
• Lambda A100: $1.10/hour (~1 min per min of video)
• Cost per 10-minute video: $0.07-0.18 (often cheaper than electricity)

Installation and Setup

Step 1: Update ComfyUI

Ensure your ComfyUI installation is current:

cd ComfyUI
git pull origin main
pip install --upgrade -r requirements.txt

Step 2: Download Upscaling Models

Download the following models:

Upscaling Models (Choose based on content):

For photorealistic videos:

File: NMKD-Superscale-SP_178000_G.pth
Type: 4x upscaler for photorealistic content
Destination: ComfyUI/models/upscale_models/
Size: 500MB
Source: https://huggingface.co/realesrgan/RealESRGAN_x4

For anime / stylized content:

File: 2x_AnimeSharper.pth
Type: 2x upscaler for anime/stylized
Destination: ComfyUI/models/upscale_models/
Size: 350MB
Source: https://huggingface.co/upscayl/upscayl-models

Face Restoration Model (Optional but highly recommended):

File: codeformer.pth
Type: Face enhancement and artifact removal
Destination: ComfyUI/models/face_restore_models/
Size: 350MB
Source: https://github.com/sczhou/CodeFormer

Download CodeFormer:

# From CodeFormer GitHub releases
wget https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth
# Move to ComfyUI face restore folder
mv codeformer.pth ComfyUI/models/face_restore_models/

VAE and Related (If not already installed):

Same models as for video generation:

• wan-2.1-vae.safetensors
• Text encoder models

Step 3: Install Custom Nodes

The upscaling workflow requires several custom nodes. Install via ComfyUI Manager:

# Launch ComfyUI
python main.py

# In web interface:
# 1. Click "Manager" button (left sidebar)
# 2. Click "Install Missing Custom Nodes"
# 3. Search for and install:
#    - Video upscale nodes
#    - RIFE frame interpolation
#    - CodeFormer face restoration
#    - Upscale model loaders

Step 4: Load the Upscaling Workflow

Download the workflow JSON file:

• File: video-upscaler-workflow.json
• Source: ComfyUI community workflows
• How to load:
  • Drag-drop into ComfyUI interface, or
  • Menu → Load → Select JSON file

Step 5: Verify Setup

Confirm all components load:

1. Upscale model loader shows available models
2. CodeFormer node appears in menu (if using face restoration)
3. RIFE node available (for frame interpolation)
4. No red startup errors

Video Upscaling Workflow Overview

Complete Pipeline Flow

Input Video File
        ↓
    [Load Video]
        ↓
    [Upscale Model Loader] ← (NMKD or Anime Sharp)
        ↓
    [Per-Frame Upscaling] → Each frame 4x resolution
        ↓
    [Optional: CodeFormer Face Restoration]
        ↓
    [Optional: RIFE Frame Interpolation] → 16 FPS to 32/60 FPS
        ↓
    [Color Correction / Grading]
        ↓
    [Final Video Export]

Key Settings Configuration

Video Input Settings

Loading Your Video:

Video File: (Select your AI-generated video)
Format: MP4, WebM, AVI accepted
Audio: Automatically preserved
Resolution: Auto-detected
Framerate: Auto-detected (usually 16 FPS for WAN videos)

Upscale Settings

Model Selection:

• For Photorealistic Videos: Use NMKD-Superscale

  • Best for: AI-generated realistic scenes, products, people
  • Quality: Excellent detail, natural appearance
  • Speed: Standard (slower than anime)

• For Anime / Stylized: Use 2x_AnimeSharper

  • Best for: Illustrative content, anime-style renders
  • Quality: Sharp linework, vibrant colors
  • Speed: Faster than photorealistic

• For Mixed Content: Use NMKD (safe default)

Upscale Multiplier:

Target Resolution: Set final output dimensions:

For 16:9 aspect ratio:
  720p: Height 720 (width auto-calculates to 1280)
  1080p: Height 1080 (width auto-calculates to 1920)
  1440p: Height 1440 (width auto-calculates to 2560)
  2160p/4K: Height 2160 (width auto-calculates to 3840)

For other aspect ratios:
  The workflow auto-preserves your original aspect ratio

Face Restoration with CodeFormer

Enabling Face Restoration:

Toggle: ON
Model: CodeFormer (default)
Fidelity: 0.5 (default)
  - 0.0 = Maximum restoration (can look over-processed)
  - 0.5 = Balanced (recommended)
  - 1.0 = Minimum restoration (keeps original character)

When to Use:

• Always use for human faces (significant improvement)
• Skip for anime/stylized (can introduce artifacts)
• Consider disabling for extreme close-ups (may over-smooth)

What CodeFormer Does:

• Removes compression artifacts
• Smooths skin imperfections
• Reduces noise
• Enhances eye clarity
• Restores detail in dark areas

Trade-offs:

• ✓ Dramatically improves face quality
• ✗ Can introduce slight temporal flickering
• ✗ May over-smooth fine details
• ✓ Solution: Use fidelity 0.7+ to minimize over-processing

Frame Interpolation with RIFE

Why Frame Interpolation?

Native AI video is 16 FPS—perceptibly choppy. Interpolation synthesizes new frames between existing ones:

Original: Frame 1 → [gap] → Frame 2 (16 FPS, choppy)
After RIFE: Frame 1 → Synthetic Frame → Frame 2 (32 FPS, smooth)

Configuration:

Enable: YES (for professional results)
Model: RIFE v4.25 (latest, best quality)
Multiplier: 2x (16 FPS → 32 FPS) or 4x (16 FPS → 64 FPS)
Quality Mode: Balanced

Recommended: 2x (to 32 FPS) balances smoothness and processing time.

Batch Processing for Low VRAM

The Challenge: Entire 60-frame video can exceed 6GB VRAM limit

The Solution: Process in batches

Enable Metab Manager: YES
Batch Size: 10
  (Process 10 frames at a time, repeat for all frames)

Actual Process:
Frames 1-10 → Upscale → Write to disk
Frames 11-20 → Clear VRAM → Upscale → Write to disk
Frames 21-30 → Continue...

Batch Size Settings:

Step-by-Step Usage Guide

Preparing Your Video

Step 1: Export from AI Generator

When exporting from WAN 2.2 or LongCat:

• Format: MP4 (H.264)
• Codec: H.264 or H.265
• Bitrate: 8-12 Mbps (quality preservation)
• Resolution: Keep native (don't pre-upscale)
• Audio: Include if available

Step 2: Import into ComfyUI

1. Load the video upscaler workflow JSON
2. Click "Load Video" node
3. Navigate to your AI-generated video
4. Select and confirm
5. Verify resolution and framerate display

Step 3: Select Upscale Model

Click "Load Upscale Model" node
Choose:
  - NMKD for photorealistic
  - AnimeSharper for anime
  - Based on your content

Step 4: Configure Settings

Resolution settings:
  - Input: Auto-detected
  - Output: Select target (1080p, 1440p, 2160p)
  - Multiplier: Set to 4x (standard) or 2x (faster)

Face restoration:
  - Toggle: ON
  - Fidelity: 0.5 (or 0.7 if heavy over-smoothing)

Frame interpolation:
  - Enable: YES
  - Multiplier: 2x (16 FPS → 32 FPS)

Step 5: Configure Batch (For Low VRAM)

If using Metab Manager:
  - Enable: YES
  - Batch Size: [Based on your GPU VRAM]

Step 6: Queue Generation

Click "Queue Prompt"
Let ComfyUI process all frames

Processing Timeline

Example: 1-minute video on RTX 4080

Upscaling (480p → 1920p): ~8 minutes
Frame Interpolation (16 → 32 FPS): ~4 minutes
Face Restoration: ~2 minutes
Total: ~14 minutes per 1-minute video

Example: 1-minute video on RTX 3060

Upscaling: ~20 minutes
Frame Interpolation: ~10 minutes
Face Restoration: ~5 minutes
Total: ~35 minutes per 1-minute video

With Batch Processing (6GB GPU):

Same upscaling: ~40 minutes (due to batch load/unload)
Frame Interpolation: ~15 minutes
Face Restoration: ~8 minutes
Total: ~63 minutes per 1-minute video

Video Export

Optimal Export Settings:

Codec: H.265 (HEVC) if possible, else H.264
Bitrate: 15-20 Mbps (1080p), 25-35 Mbps (4K)
Framerate: 30 FPS (standard) or 60 FPS (smooth)
Audio: AAC, 128 kbps
Container: MP4

Advanced Techniques

Selective Face Restoration

Problem: Face restoration can over-smooth fine details or introduce flickering

Solution: Dual-Pass Approach

Pass 1: Upscale all frames WITHOUT face restoration
Pass 2: Extract face regions only
Pass 3: Apply CodeFormer only to extracted faces
Pass 4: Composite enhanced faces back into upscaled frames

This keeps environments sharp while enhancing faces.

Multi-Pass Enhancement

For Maximum Quality:

Pass 1: Upscale 4x (480p → 1920p)
Pass 2: Upscale 1.5x more (1920p → 2880p / 3K)
Pass 3: Face restoration on final resolution
Pass 4: Frame interpolation for final smoothness

Result: 3K video at 32 FPS with pristine faces

Trade-off: 2-3x processing time, exceptional quality.

Custom Model Combinations

For Anime + Photorealistic Mix:

Scenes 1-5 (anime): Use AnimeSharper
Scenes 6-10 (realistic): Switch to NMKD
Composite together in editing software

Or: Use NMKD for all (safe default, reasonable for both)

Video-Specific Optimizations

For Face-Focused Content (talking head, interview):

High face restoration (fidelity 0.3-0.5)
4x upscale (maximum face detail)
2x frame interpolation
Result: Interview-quality footage

For Wide Shots / Landscape:

Lower face restoration (fidelity 0.7+)
4x upscale
2x frame interpolation
Result: Environmental detail preserved

For Fast-Moving Action:

Moderate face restoration (fidelity 0.6)
4x upscale
4x frame interpolation (smoother motion)
Result: Cinematic action with smooth motion

Troubleshooting

Issue 1: Processing Out of Memory

Symptoms: "CUDA out of memory" error mid-upscaling

Solutions:

1. Enable batch processing (if not already)
2. Reduce batch size: 20 → 10 → 5
3. Lower target resolution: 2160p → 1440p
4. Disable frame interpolation temporarily
5. Process in chunks (first 30s, then last 30s)

Issue 2: Face Looks Over-Smoothed

Symptoms: Faces look plastic or unnatural

Solutions:

1. Lower face restoration fidelity: 0.5 → 0.7 → 1.0
2. Disable CodeFormer entirely (fidelity = 1.0)
3. Use dual-pass approach (restore only face regions)
4. Reduce upscale factor: 4x → 2x

Issue 3: Frame Interpolation Introduces Artifacts

Symptoms: Ghosting, motion blur, or blending artifacts between frames

Solutions:

1. Disable frame interpolation (use native 16 FPS)
2. Try lower multiplier: 4x → 2x
3. Use different RIFE model (if available)
4. Accept minor artifacts (often acceptable in final video)

Issue 4: Processing Takes Extremely Long

Symptoms: Upscaling 1-minute video takes 2+ hours

Causes:

• Batch size too small (excessive VRAM swaps)
• Target resolution too high (2160p+ on limited VRAM)
• All enhancements enabled on low-end GPU
• Frame interpolation with high multiplier (4x)

Solutions:

1. Reduce batch size gradually until processing is faster
2. Lower target resolution: 2160p → 1440p
3. Disable less critical features (face restoration, interpolation)
4. Use cloud GPU for one-time projects

Issue 5: Audio Desynchronized After Export

Symptoms: Audio drifts from video over time

Cause: Frame interpolation changes video framerate without adjusting audio

Solution:

If interpolating from 16 FPS → 32 FPS:
  Video becomes 2x longer
  Must re-sync audio or speed up audio 2x
  Use video editor: Separate audio, time-stretch to match

Comparison: Upscaling Tools

Real-World Example: Professional Product Video

Scenario: Generate product video at 480p, upscale to cinema-quality

Original Video Stats

• Resolution: 852×480
• Framerate: 16 FPS
• Duration: 30 seconds
• Content: Product rotating, showing features

Upscaling Pipeline

Input: 30-second 480p WAN 2.2 generated product video

Configuration:

Upscale Model: NMKD-Superscale (photorealistic)
Upscale Factor: 4x (480p → 1920p)
Face Restoration: OFF (no faces in product video)
Frame Interpolation: 2x (16 FPS → 32 FPS)
Target Resolution: 1920×1080 (1080p)
Batch Size: 20 frames

Processing Timeline (RTX 4080):

Upscaling 480 frames: 6 minutes
Frame interpolation (480 → 960 frames): 4 minutes
Export and encode: 2 minutes
Total: 12 minutes

Output Video:

• Resolution: 1920×1080 (Full HD)
• Framerate: 32 FPS (smooth)
• Duration: 30 seconds
• Quality: Professional, broadcast-ready

Final Steps:

1. Import into DaVinci Resolve
2. Add: Product name overlay, features text
3. Add: Background music (royalty-free)
4. Color grade: Warm, professional tones
5. Export: 1080p, 30 FPS, H.265 codec

Result: 30-second professional product video from AI generation, suitable for website, YouTube, sales presentations.

Performance Benchmarks

Real Processing Times

System 1: RTX 3060 (12GB VRAM)

Test: 60-frame video (3.75 seconds @ 16 FPS)
Task: 4x upscale + CodeFormer + 2x interpolation

Upscaling: 3.2 minutes
Face Restoration: 0.8 minutes
Interpolation: 2.0 minutes
Total: 6.0 minutes for 3.75 seconds of video
Extrapolated: ~96 minutes per 60-second video

System 2: RTX 4080 (16GB VRAM)

Test: Same 60-frame video

Upscaling: 1.2 minutes
Face Restoration: 0.3 minutes
Interpolation: 0.8 minutes
Total: 2.3 minutes for 3.75 seconds
Extrapolated: ~37 minutes per 60-second video

System 3: RTX 4090 (24GB VRAM)

Test: Same 60-frame video

Upscaling: 0.6 minutes
Face Restoration: 0.15 minutes
Interpolation: 0.4 minutes
Total: 1.15 minutes for 3.75 seconds
Extrapolated: ~18 minutes per 60-second video

Conclusion

Video upscaling transforms rough AI-generated footage into professional-grade deliverables. The ComfyUI workflow provides industrial-strength capabilities accessible to creators with any GPU, from budget 6GB cards to premium RTX 4090s.

Start with standard settings: 4x upscaling, CodeFormer face restoration, 2x frame interpolation. Most videos improve dramatically. Experiment with settings for specific content types—face-heavy content benefits from aggressive restoration, while environments benefit from lower fidelity to preserve detail.

The investment in upscaling pays dividends. Videos that looked "pretty good for AI" suddenly look professionally shot. This bridges the final gap between prototype and production-ready content.

Related Resources

1. LongCat Video Complete Guide - Generate high-quality AI videos before upscaling
2. LongCat FP8 Workflow: Memory-Efficient Video - Create videos on budget hardware
3. ComfyUI LongCat Workflow for Low VRAM - Optimize workflows for limited GPU memory