Tech 5 min read

Breaking Down ComfyUI's Official Complete AI Upscaling Handbook

IkesanContents

ComfyUI published a comprehensive AI upscaling guide as the second most requested feature from their community. After weeks of testing and integration, all major upscaling methods are now available on ComfyUI.

The guide covers 10 use cases and 20 workflows — pretty practical stuff, so here’s a summary.

Original article: The Complete AI Upscaling Handbook: All in ComfyUI

Why Upscaling Now

Reasons the guide highlights:

  • Industry requirements: 4K/8K deliverables are becoming standard in VFX, film, marketing, e-commerce, gaming, and design
  • Reframing: When changing aspect ratios or cropping, insufficient resolution means quality loss
  • AI generation quality issues: Most AI-generated images are still 480-720p — edges, textures, faces, and hands aren’t production-quality at the pixel level
  • Cost and time efficiency: Generating small then upscaling at the end is the go-to strategy for budget- and time-constrained pipelines

Personally, I’d been generating at whatever size fits on my machine, so I never felt the need for upscaling. But the “generate → fix → upscale → deliver” pipeline concept makes sense, as does adding texture via creative upscaling. Especially for video, the “generate small and fast → enlarge later” approach seems to pay off.

Conservative vs Creative Upscaling

The most important concept in this guide is distinguishing these two approaches.

Conservative upscaling faithfully increases resolution while staying true to the original image. Similar to traditional super-resolution, it’s suited for product photos, architecture, and other “must not alter” use cases.

Creative upscaling has AI fill in and reconstruct details that didn’t exist. Made possible by diffusion models, it’s for “make it better” use cases like adding skin texture or enhancing landscape atmosphere.

Neither is superior — the key is choosing based on your use case.

Image Upscaling: Use Case Guide

Portraits and Skin Texture

Adding realistic skin detail (pores, natural skin texture) while maintaining character consistency is critical. In this area, Magnific Image Skin Enhancer reportedly dominates. It specializes in making the plastic-looking skin typical of AI-generated images look natural.

Product Photography

Faithful reproduction of material textures, label edges, and small text requires conservative models. Recommended: HitPaw, Magnific Precise, WaveSpeed SeedVR2, Nano Banana Pro.

Landscapes and Environments

Depends on your needs. Creative models for boosting atmosphere and detail, conservative models for preserving building accuracy. If the input has artifacts, creative models might reconstruct them, but conservative models will leave them as-is.

Illustrations and Stylized Art

For inputs with a distinctive style, choose conservative models. Creative models can add too much detail and break the style. Conversely, for inputs with room for more detail, creative models can enhance the style. Adjusting the “creativity” parameter in Magnific Creative or Topaz Image Enhance works well.

Fixing AI-Generated Images

Using upscaling to fix typical AI image problems — extra fingers, artifacts, anatomical inaccuracies — is not recommended. Creative upscaling sometimes fixes them, often not. Best practice is to fix issues with image editing models or conventional tools before upscaling.

Image Upscaling Speed (1K → 4K)

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

Video Upscaling: Use Case Guide

Quality Mode

SeedVR2 and HitPaw produce the most accurate results. High image quality with excellent character consistency. Topaz is also effective but tends toward the creative side. It’s an option if a stylized cinematic look is acceptable.

Speed Mode

FlashVSR excels when speed is the priority. 720p → 1080p takes about 41 seconds for a 10-second clip. HitPaw is also relatively fast. Topaz has nearly identical rendering speeds between Fast and Creative modes, so using Creative mode without sacrificing speed is recommended.

Fixing AI Video

Topaz Astra Creative generates very smooth, visually consistent output. Strong for AI-generated video restoration. Workflow tip: run a fix pass first, upscale to 1080p, then progressively upscale to 4K.

Archive Restoration

SeedVR2 is particularly impressive. It preserves the original footage’s structure and detail while making it clearer. Avoids adding artificial, excessive features.

CGI and 3D Renders

Topaz, SeedVR2, and HitPaw all perform well. They produce clean, sharp results while maintaining edge crispness — critical for composited imagery.

Video Upscaling Speed (10s 720p clip)

ModelResolutionProcessing Time
FlashVSR (local)1080p~41s
FlashVSR (local)4K~52s
HitPaw2K~80s
HitPaw4K~175s
SeedVR2 (local)1080p~312s
Topaz Astra1080p~374s
Topaz Astra4K~560s
Wan2.2 Upscale (local)~1100s (2s input)

Hardware: NVIDIA GeForce RTX 5090

Available Workflows in ComfyUI

All models and workflows are immediately available on Comfy Cloud. They can also be downloaded and set up locally.

Image Upscaling:

  • SeedVR2 (open source)
  • Topaz Image Enhance: Creative / Landscape / Illustration
  • Magnific: Skin Enhancer
  • HitPaw: Creative / Creative Portrait
  • Recraft: Creative / Crisp
  • Nano Banana Pro: Portrait / Product / Styled Art
  • Traditional non-AI upscaling (open source)

Video Upscaling:

  • Fast GAN (open source): Traditional
  • SeedVR2 (open source / WaveSpeed API): Conservative
  • FlashVSR (WaveSpeed API): Conservative
  • Wan2.2 Upscale (open source): Creative
  • Topaz Astra: Creative
  • HitPaw: All-in-one