ComfyUI Image Upscaling: Workflows for High-Resolution AI Art

Standard AI image generation produces moderate resolutions that may need upscaling for final use. ComfyUI offers multiple upscaling approaches, from simple model-based upscaling to sophisticated multi-pass techniques that enhance detail while increasing resolution.

This guide covers upscaling workflows from basic to advanced, helping you produce high-resolution output for any use case.

Quick Answer: ComfyUI upscaling ranges from simple upscaler models (ESRGAN, Real-ESRGAN) to complex multi-pass techniques (Ultimate SD Upscale, tiled VAE). For quick 2x upscaling, use a quality upscaler model. For maximum quality at higher scales, combine upscaler models with img2img refinement passes. VRAM limitations require tiled approaches for large images.

:::tip[Key Takeaways]

• ComfyUI Image Upscaling: Workflows for High-Resolution AI Art represents an important development in its field
• Multiple approaches exist depending on your goals
• Staying informed helps you make better decisions
• Hands-on experience is the best way to learn :::

Understanding Upscaling Options

Upscaling Approaches

Model-based upscaling: Neural network upscalers trained to add detail while scaling.

Latent upscaling: Scaling in latent space with regeneration.

Tiled upscaling: Processing sections for VRAM efficiency.

Multi-pass: Combining techniques for maximum quality.

Quality Considerations

Upscaling trade-offs:

Resolution vs detail: Higher scale can introduce artifacts.

Speed vs quality: Better results take more time.

VRAM vs size: Larger images need more memory.

Preservation vs enhancement: Keeping original vs adding detail.

Upscaler Models

Popular Upscaler Types

ESRGAN/Real-ESRGAN:

• Good general-purpose upscaling
• Various trained variants
• Balanced quality/speed

4x-UltraSharp:

• Excellent detail enhancement
• Good for photorealistic content
• Slightly slower

Remacri/Nomos:

• Community-developed options
• Various optimization targets
• Worth testing for your use case

Swin2SR:

• More recent architecture
• Good quality results
• Higher VRAM usage

Model Selection

Choose based on content:

Photorealistic: 4x-UltraSharp, Real-ESRGAN Anime/Illustration: 4x-AnimeSharp, specialized anime upscalers General purpose: Real-ESRGAN, Remacri

Test multiple options on your specific content.

Installing Upscalers

Place upscaler models in appropriate folder:

ComfyUI/models/upscale_models/

Download from sources like OpenModelDB or Civitai.

Simple Upscaling Workflow

Basic Upscaler Node

The simplest approach:

[Load Image]
      ↓
[Load Upscale Model]
      ↓
[Upscale Image (Using Model)]
      ↓
[Save Image]

This applies the upscaler directly without additional processing.

Settings

Model selection: Choose appropriate upscaler.

Scale factor: Usually 2x or 4x depending on model.

When to Use

Simple upscaling works for:

• Quick previews
• Already high-quality source images
• When speed matters most
• Minor resolution increases

Latent Upscaling

Upscaling in Latent Space

Scale before decoding:

[Generate Image (KSampler)]
        ↓
[Upscale Latent]
        ↓
[KSampler] ← Low denoise for refinement
        ↓
[VAE Decode]
        ↓
[Save Image]

Benefits

Detail addition: Regeneration adds fine details.

Style consistency: Uses same model for enhancement.

Quality potential: Can produce better results than pure upscaling.

Denoise Settings

For latent upscale refinement:

0.2-0.4: Light detail enhancement, preserves original closely.

0.4-0.6: Moderate enhancement, more detail addition.

0.6+: Significant changes, may alter composition.

Start low and increase if needed.

Multi-Pass Upscaling

Ultimate SD Upscale

Combining upscaler with img2img passes:

First pass: Upscale to intermediate resolution.

Second pass: Img2img refinement at full resolution.

Optional third: Final polish pass.

Implementation

[Load Image]
      ↓
[Upscale Image (Using Model)] ← 2x
      ↓
[VAE Encode]
      ↓
[KSampler] ← denoise 0.3-0.5
      ↓
[VAE Decode]
      ↓
[Upscale Image (Using Model)] ← 2x again
      ↓
[VAE Encode]
      ↓
[KSampler] ← denoise 0.2-0.3
      ↓
[VAE Decode]
      ↓
[Save Image]

Quality Gains

Multi-pass produces:

• More coherent detail addition
• Better texture quality
• Reduced upscaling artifacts
• Higher maximum quality

Trade-off is significantly longer processing time.

Tiled Upscaling

Why Tiling

Large images exceed VRAM:

• 2048x2048 at 8GB may struggle
• 4096x4096 needs significant VRAM
• Tiling processes sections sequentially

Tiled VAE

For decoding large images:

Tiled VAE Decode: Process VAE in tiles.

Benefits: Enables large images on limited VRAM.

Trade-offs: Slightly longer, potential tile boundaries.

Tiled Diffusion

Ultimate SD Upscale with tiles:

• Process upscale in overlapping sections
• Blend between tiles
• Enables very large outputs

Tile Settings

Tile size: 512 or 768 typical.

Overlap: 64-128 pixels for smooth blending.

Batch size: Adjust based on VRAM.

Workflow Optimization

Choosing Approach by Use Case

Quick preview: Simple upscaler only.

Final output: Multi-pass with refinement.

Very large: Tiled approach required.

Speed priority: Single pass with good upscaler.

Quality priority: Multiple passes with lower denoise.

Resolution Planning

Plan your target:

Source: Typical 512 or 1024 from generation.

Target: What final size do you need?

Steps: 2x or 4x per pass.

Iterations: How many enhancement passes.

VRAM Management

Handle memory efficiently:

• Enable tiled processing if needed
• Process one image at a time for large outputs
• Use lower precision if struggling
• Close other applications

Common Issues and Solutions

Blurry Results

Causes:

• Weak upscaler model
• Too aggressive denoise reducing detail
• Source image too low quality

Solutions:

• Try different upscaler
• Reduce denoise in refinement
• Start with better quality generation

Visible Tile Boundaries

Causes:

• Insufficient overlap
• Inconsistent processing between tiles
• Denoise too high in tiled mode

Solutions:

• Increase tile overlap
• Use consistent settings
• Lower denoise for tiled processing

Detail Hallucination

Causes:

• Denoise too high
• Too many refinement passes
• Model adding inappropriate detail

Solutions:

• Reduce denoise settings
• Fewer passes
• Different upscaler model

Memory Errors

Causes:

• Image too large for VRAM
• Too many simultaneous operations

Solutions:

• Enable tiled processing
• Reduce tile size
• Lower batch size

Quality Assessment

Checking Results

Evaluate upscaled images:

Zoom inspection: Look at 100% zoom.

Compare to source: Check preservation.

Detail quality: Are added details appropriate?

Artifact check: Look for upscaling problems.

Common Artifacts

What to watch for:

Oversharpening: Halos around edges.

Texture repetition: Unnatural patterns.

Blur patches: Areas lacking detail.

Color shifts: Changed colors from original.

Frequently Asked Questions

What's the best upscaler for photorealistic images?

4x-UltraSharp and Real-ESRGAN work well. Test both on your specific content.

How much can I upscale?

2x-4x per pass is typical. Beyond 4x typically needs multiple passes.

Does upscaling add real detail?

Sort of. AI upscalers add plausible detail based on training. It's generated, not recovered.

Why does upscaling take so long?

Multi-pass and tiled approaches process images multiple times. Quality requires time.

Can I upscale any image in ComfyUI?

Yes, but results depend on source quality. Better source = better upscale.

What tile size should I use?

512 is safe for most VRAM. 768 or higher if you have overhead.

Should I upscale during generation or after?

After is more flexible. Generation at higher resolution is also option if VRAM allows.

How do I handle images with both photorealistic and stylized elements?

General purpose upscalers work best. Or upscale regions separately.

Conclusion

ComfyUI upscaling ranges from simple one-step processes to complex multi-pass workflows. Your choice depends on quality requirements, available VRAM, and time constraints.

For most use cases, start with a quality upscaler model like Real-ESRGAN for quick results. When quality matters most, add img2img refinement passes with low denoise. For large outputs, enable tiled processing to work within VRAM limits.

For character-specific upscaling that maintains consistency, see our character consistency guide. For basic ComfyUI orientation, check our essentials guide.