Skip to main content
This guide covers image upscaling workflows in ComfyUI, including local models and partner node options for various use cases.

The Complete AI Upscaling Handbook

For a comprehensive deep-dive with benchmarks and visual comparisons, read the full article on the ComfyUI Blog.

Why upscaling matters

  • Industry requirements for resolutions: 4K/8K end results are widely required across VFX & Film, Marketing, E-commerce, Gaming, and Design industries, and requirements keep moving up.
  • Reframing: When assets ship across different aspect ratios and placements, extra resolution headroom prevents quality loss after cropping or recomposition.
  • The AI content problem: Generative AI has dramatically increased the speed of visual creation, but most AI assets today are still generated at 480–720p, and not production-ready at the pixel level: edges, micro-texture, faces/hands, compression artifacts, etc. This creates a new pipeline: Generation → Refinement/Fix → Upscale → Delivery.
  • Cost/time efficiency: Generate or render smaller → upscale at the end. This is a standard efficiency move in pipelines under budget/time pressure.

Key concepts

Upscaling vs. enhancement

  • Upscaling increases resolution while reconstructing detail.
  • Enhancement improves perceived quality (denoise, sharpen, restore, color, faces, etc.).
The best practice in AI pipelines usually combines both steps.

Creative vs. conservative upscale

Diffusion models and generative AI changed what “upscaling” means. Traditionally, super-resolution aimed to preserve the original signal. Today, some models can re-imagine details that never existed. And there are models approaching different balances between the two.
ConservativeCreative
ApproachPreserve the originalReimagine and enhance
StrengthsHigh accuracy, consistency, production-safeRich details, sharper, more visually striking
LimitsCan appear flat, limited improvement on poor inputPossible hallucinations or structural drift
ModelsMagnific Precise, SeedVR2, FlashVSR, Topaz Fast, HitPawWan 2.2, Magnific Creative, Topaz Astra, HitPaw Creative

Use cases

TL;DR
  • Portraits: Magnific Skin Enhancer
  • Product photography: Magnific Precise, WaveSpeed SeedVR2, or Nano Banana Pro
  • Landscapes & illustrations: Model choice depends on your specific needs
  • AI artifacts: Do not rely on upscaling to fix common AI artifacts
  • SeedVR2 tip: Downscale the image to 0.35 megapixels with the ImageScaleToTotalPixels node before upscaling for better results

Portrait and skin enhancement

When it comes to upscaling portraits of realistic subjects, achieving realistic skin details while maintaining character consistency is key. The best upscale models for fixing plastic skin need to add texture, pores and natural skin imperfections. In this domain, Magnific Image Skin Enhancer far outperforms the rest. Recommended: Magnific Image Skin Enhancer

Product photography

Maintaining a faithful representation of materials, product label edges, and small text is a must when upscaling product images. Therefore, conservative upscaling models are necessary. Recommended: HitPaw, Magnific Precise, WaveSpeed SeedVR2, Nano Banana Pro

Landscape and environments

For this use case, the upscale model you select really depends on your needs. Have an environmental shot that you want to pop with details and set a mood? Use a creative upscale model. Have an establishing shot where a building needs to stay consistent? Use a conservative upscale model. It’s worth noting that if your input image has artifacts, creative upscale models may be able to reimagine the artifact while conservative upscale models will not.

Stylized art and illustration

This is another use case where the correct upscale model depends on your needs. A rule of thumb is that if your input has a very distinct/unique style, the best model will be conservative. Creative models may add too many details and stray away from the desired illustrated style. However, if your input image has room for detail, a creative model works well and can even improve the style. We recommend playing around with the ‘creativity’ parameters for Magnific Creative and Topaz Image Enhance to find values that meet your needs! Nano Banana Pro can also work well for more common styles and adding details (but may rely on ‘seed luck’). Recommended: Magnific Creative, Topaz Image Enhance, Nano Banana Pro

AI-generated images

When it comes to upscaling images that have typical AI issues such as too many fingers, artifacts, incorrect anatomy, and morphing, the assumption is that creative upscales can fix it. Sometimes that’s true, but other times not so much. Best practice is to resolve these issues prior to upscaling with image-edit models or traditional tools (or generate a new image entirely).

Available models

Local models (ESRGAN)

For basic local upscaling using ESRGAN models, see the basic upscale tutorial.
ModelBest for
RealESRGANGeneral-purpose upscaling
BSRGANText and sharp edges
SwinIRNatural textures, landscapes

General creative upscaling

For general creative upscaling across various use cases:

Partner nodes

Partner nodes provide access to advanced upscaling models via API.
ModelTypeFeatures
Topaz Image Enhance (Bloom)CreativeSubject detection, face enhancement, color preservation, up to 8K
Magnific PreciseConservativeHigh accuracy, production-safe
Magnific CreativeCreativeDetail reimagination
Magnific Skin EnhancerCreativePortrait-specific, adds realistic skin texture
Nano Banana ProConservativeFast, good for product photography
WaveSpeed SeedVR2ConservativeHigh fidelity
HitPawBothConservative and creative modes
RecraftBothCreative and crisp modes

Benchmark: 1K to 4K upscale time

ModelTime
Magnific Precise~40s
WaveSpeed SeedVR2~40s
Magnific Creative~50s
Magnific Skin Enhancer~60s
HitPaw~60s
Nano Banana Pro~80s
Topaz Image Enhance~100s

Tips

  • For SeedVR2, downscale the image to 0.35 megapixels with the ImageScaleToTotalPixels node before upscaling for better results.
  • Chain multiple upscale nodes (e.g., 2x → 4x) for ultra-high magnification.
  • Connect upscale nodes after generation for “generate + enhance” pipelines.
  • Test multiple models on your specific image type to find the best fit.