Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

Prologue — Small to Big, Blurry to Sharp

You have probably hit this wall: an 800x600 photo from a family album looks pixelated on a 4K TV. An e-commerce seller wants to scale a 1024px product shot to 4000px for print. A video editor wants to master old SD footage at 4K.

Classic bicubic or Lanczos interpolation does not invent information — it just smoothly stretches what already exists. **AI upscalers** are different. A neural network trained on millions of high-res/low-res pairs **synthesizes plausible detail**.

In 2026 the space has exploded — commercial Topaz Photo AI v3, Magnific, Krea on one side; open-source Real-ESRGAN, SUPIR, CodeFormer on the other; ComfyUI workflows bridging them.

This post charts the whole map.

Chapter 1 — Why Image Upscaling Matters

Scenarios that need upscaling:

- **Old family photos** — restore noise and blur from 1990s digital cameras (VGA, 0.3MP) or scanned film

- **Low-resolution inputs** — mobile screenshots, tiny thumbnails downloaded from the web, old game captures

- **Print scaling** — an A4 print at 300dpi needs roughly 2480x3508px, but the source is smaller

- **Video still extraction** — capture a frame from 1080p footage and turn it into a poster

- **E-commerce and real estate** — blow up a product thumbnail to banner size

- **Manga and illustration remasters** — bring 1990s and 2000s low-res art to 4K displays

The point is not just enlargement. **Denoising, compression-artifact removal, color correction, face detail enhancement** all running in one pipeline is the 2026 standard.

Chapter 2 — Commercial AI Upscalers

Topaz Photo AI v3

Topaz Labs flagship. It merges the previously separate Gigapixel (upscale), Sharpen, and DeNoise products into one UI.

- **Price**: roughly 199 dollars one-time (perpetual license + 1 year of updates)

- **Models**: High Fidelity, Standard, Low Resolution, CGI, Lines, Art and CG, Recover — case-specific options

- **Strengths**: clean GUI, batch processing, local GPU acceleration, Photoshop plug-in

- **Weaknesses**: more conservative detail synthesis than diffusion-based SUPIR — portraits can still feel a touch "AI-ish"

Topaz Gigapixel 8

Standalone upscale tool. Versions 8.x have added **Recovery v2** and **Redefine** as diffusion-strength modes.

- **Recovery mode** — restores faces and text even at very low input resolution

- **Redefine** — diffusion-based "reimagining" mode with high creative freedom

Magnific

Diffusion-based commercial upscaler that exploded on social media after launching in 2024. Positioning: "do not just enlarge an image, **reinterpret** it".

- **Magnific Sharpen** (2024) — sharpening-focused product

- The **creativity slider** is the headline knob — 0 stays conservative, 10 paints in entirely new detail

- **HDR** and **Resemblance** parameters tune tonal range and fidelity to the source

- Subscription pricing (Pro around 39 dollars per month, Premium up to 99 dollars)

Krea Upscaler / Krea Enhancer

The upscale function inside the Krea AI platform. **Faster and lighter than Magnific** is the selling point. Live previews and tight workflow integration.

Adobe Photoshop Super Resolution

The Camera RAW **Enhance** feature. An ML model that doubles pixel count (4x area). Free for Creative Cloud subscribers and very consistent.

Luminar Neo Upscale AI

Skylum extension to the Luminar Neo editor. Pairs well with the same vendor automatic landscape and portrait corrections.

ON1 Resize AI

The AI version of ON1 Resize, long used in print. Detail preservation at large output sizes is the strength.

Pixelmator Pro ML Super Resolution

Apple Silicon native. For macOS users this is the fastest no-extra-cost AI upscaler in the basic app price. Core ML accelerated.

Bigjpg / Lets Enhance

Web-based services. No install — upload in browser and download. Fine for light use; inefficient for large or batch jobs.

Chapter 3 — Open Source Super Resolution Models

Real-ESRGAN

Effectively the standard, from Xintao Wang at Tencent ARC. Trained with **practical degradation** (synthesized real-world noise, JPEG, blur), so it generalizes well to real photos.

- **Variants**: RealESRGAN_x4plus (general), RealESRGAN_x4plus_anime_6B (manga), RealESRNet (less noise)

- ncnn Vulkan backend runs on CPU, GPU, and Apple Silicon

- Weights are BSD 3-Clause (commercial use allowed; always check the model card)

Real-CUGAN

A Bilibili-released cousin specialized for manga and anime. Cleaner line preservation for illustration.

SwinIR

Swin Transformer-based super resolution. Very high PSNR and SSIM. **Better when you want accurate pixel restoration over invented detail.**

HAT (Hybrid Attention Transformer)

Chen et al, CVPR 2023. A channel-and-spatial attention combo that surpasses SwinIR. Recent academic baseline.

DAT (Dual Aggregation Transformer)

ICCV 2023. Combines attention along both axes (spatial and channel) for efficiency.

DRCT (Dense-Residual-Connected Transformer)

CVPRW 2024. Tackles SwinIR known information-flow loss with dense connections.

CCSR (Controllable Conditional Super-Resolution)

Slider-controllable generative freedom. Often paired with SUPIR-family ComfyUI workflows.

Chapter 4 — Diffusion-Based Restoration — Synthesizing Real Detail

SUPIR (Scaling-UP Image Restoration)

Yu, Lin, Zhang et al, ECCV 2024. Essentially **Stable Diffusion XL as a backbone for restoration**. You can pass a text prompt for semantic guidance — the model knows whether it is restoring a face, a landscape, or text.

- Heavy VRAM (24GB recommended, 16GB minimum)

- Top open-source quality as of 2026

- Well-integrated into ComfyUI, widely used in practice

StableSR

Wang et al, IJCV 2024. Uses Stable Diffusion as a super-resolution prior. Lighter than SUPIR with slightly less detail.

ResShift

Yue Z. et al, NeurIPS 2023. An efficiency-first model that slashes diffusion steps (often 15 to 20) for speed.

SeeSR

Wu et al, CVPR 2024. **Semantic-aware** diffusion SR. Injects semantic prompts during training to improve object-level fidelity.

DiffBIR

Lin et al, ECCV 2024. **Blind Image Restoration** — restoration without knowing the type of degradation. Two-stage pipeline (IR module plus diffusion refinement).

PASD (Pixel-Aware Stable Diffusion)

Yang et al, ECCV 2024. Strengthens pixel-level alignment. Especially strong on text-heavy images such as documents and signage.

Topaz Bloom AI

Topaz 2025 diffusion-based module. The commercial answer to SUPIR and Magnific.

Chapter 5 — Models Specialized for Face Restoration

Distinct from general upscaling, a separate family of models tackles **faces** only. Faces are where our eyes are most sensitive, so specialized models outperform.

GFPGAN

Xintao Wang et al, CVPR 2021. Same team as Real-ESRGAN. Realistically restores very blurry or damaged faces. Almost a standard for old-photo restoration.

- Model versions: v1.3, v1.4 (most used)

- GAN architecture using a StyleGAN2 prior

CodeFormer

Shangchen Zhou et al, NeurIPS 2022. Codebook learning that lets you slide between identity preservation and fidelity (the w parameter, 0 to 1).

- **w=0**: fidelity first (close to source, less detail)

- **w=1**: quality first (smooth, but face may shift slightly)

- Strong for old photos, surveillance frames, and low-res video conferencing captures

RestoreFormer / RestoreFormer++

Wang et al, CVPR 2022. Multi-scale cross-attention for face detail restoration.

GPEN

Yang et al, CVPR 2021. GAN Prior Embedded Network. Works even on heavily degraded inputs.

Practical tip: the **whole image is upscaled with Real-ESRGAN, then faces are re-processed with CodeFormer or GFPGAN** as a two-stage pipeline. This pattern appears in ComfyUI workflows, the Stable Diffusion WebUI Extras tab, and the Topaz Photo AI Face Recovery toggle.

Chapter 6 — Anime and Illustration Specialists

Photos and illustrations follow different statistics — clean lines, flat color fields, almost no noise. Dedicated models are needed.

waifu2x

A 2015 classic. Simple CNN, still useful for manga noise and detail. ncnn Vulkan builds cover macOS, Windows, Linux.

Real-ESRGAN Anime (RealESRGAN_x4plus_anime_6B)

The manga-tuned Real-ESRGAN variant. Effectively the 2026 standard for manga upscaling.

Real-CUGAN

Released by Bilibili. Very good color and line preservation for illustrations.

Anime4K

Real-time GPU shader. Lives inside video players (mpv, MPC-HC) and **upscales during playback**. Lower quality than model-based options but the cost is zero.

Chapter 7 — Video Upscalers

Video adds the constraint of temporal consistency — if the model output drifts frame to frame, you get flicker.

Topaz Video AI

Formerly Video Enhance AI. Video upscale plus deinterlace plus denoise plus frame interpolation in one tool. **Proteus**, **Artemis**, **Iris**, **Rhea** are model classes.

- Outputs up to 8K

- Price: roughly 299 dollars one-time

- NVIDIA and Apple Silicon accelerated

NVIDIA RTX VSR (Video Super Resolution)

Driver-level real-time upscaling of **browser video playback**. Auto-applies in Chrome and Edge. Requires RTX 30, 40, or 50 series.

Microsoft Auto SR

Introduced in 2024 on Snapdragon X Copilot+ PCs. OS-level upscaling combined with HDR tone mapping. Integrated into Windows 11.

AMD FidelityFX Super Resolution (FSR) / NVIDIA DLSS

Gaming SR. The engine renders at lower resolution and SR fills in — higher frame rate. DLSS is RTX-only; FSR runs on most GPUs.

Real-ESRGAN ncnn Vulkan + Video Scripts

The open-source workflow: split into frames, upscale each, re-encode. ffmpeg plus realesrgan-ncnn-vulkan. Free but slow.

FlowFrames

Frame interpolation tool (24fps to 60fps). Combined with upscaling, old footage starts to feel cinematic.

DaVinci Resolve Super Scale + AI Magic Mask

A built-in feature of Resolve Studio. Natural extension if you already edit in that UI.

Chapter 8 — Phone and Camera Built-in AI Touch-up

Samsung Galaxy AI Photo Edit

Galaxy S24 and S25 line. The Photo Editor AI eraser, expand, and resolution boost. Mixed cloud and on-device processing.

Apple Photo Cleanup + Image Playground

Since iOS 18, the Photos app Cleanup feature removes unwanted objects. Image Playground is generative but ties into image enhancement.

Google Pixel Magic Eraser / Best Take / Magic Editor

Pixel 8 and 9 line. The Magic Editor Reimagine option redraws selected regions via diffusion — effectively a mini restore/upscale.

What these built-ins share: **fast and natural results, limited control**. Pro work still wants Topaz, Magnific, or SUPIR.

Chapter 9 — Old Photo Restoration — End to End Pipelines

Bringing Old Photos Back to Life (Microsoft)

Wan et al, CVPR 2020. Integrated pipeline that handles fading, scratches, and paper damage. Open source (MIT).

CodeFormer-Based Pipelines

Turning on `face_restore=True, background_enhance=True, face_upsample=True` runs face restoration and background upscaling in one pass. The most-used recipe for old photos.

GFPGAN + Colorization

Black-and-white photos: GFPGAN to restore faces, then DeOldify or similar to colorize. Two-stage and common.

Photoshop Neural Filters

Adobe Sensei ML filters — **Photo Restoration**, **Colorize**, **Smart Portrait**. Easy UI, instant feedback.

Recommended Real Workflow

1. Scan at 600dpi or higher (better input means better output)

2. Photoshop Neural Filter Photo Restoration to remove scratches and dust

3. CodeFormer or GFPGAN for faces

4. SUPIR or Topaz Photo AI for the whole image

5. DeOldify if you want to colorize

6. Tonal and color finishing

Chapter 10 — ComfyUI Workflows — the Open Source Practical Platform

In 2026 the de facto open-source GUI is ComfyUI. A node graph wires together SUPIR, CCSR, Real-ESRGAN, and GFPGAN.

A representative workflow:

LoadImage

-> ImageScale (initial interpolation)

-> SUPIR Sampler (diffusion restoration, 4x)

-> Face Restore (CodeFormer)

-> SaveImage

Pros: **fully free, every parameter exposed, batch processing, community-shared JSON workflows**.

Cons: a learning curve. Finding a workflow that suits your case takes iteration.

Stable Diffusion WebUI / Forge

The A1111 or Forge UI Extras tab gives slider-based Real-ESRGAN and GFPGAN. The fastest on-ramp.

Chapter 11 — Evaluation Metrics — How to Measure "Good"

Quantitative metrics that compare upscale outputs:

- **PSNR (Peak Signal-to-Noise Ratio)** — pixel difference. Higher is closer to source. But **only weakly correlated with perceived quality.**

- **SSIM (Structural Similarity)** — structural similarity. Better aligned with human perception than PSNR.

- **LPIPS (Learned Perceptual Image Patch Similarity)** — feature distance from a pretrained network. Tracks human perception well. Lower is better.

- **FID (Frechet Inception Distance)** — for generative model evaluation. Distribution distance.

- **MUSIQ (Multi-scale Image Quality Transformer)** — no-reference quality (no source needed).

- **NIQE (Natural Image Quality Evaluator)** — no-reference statistical naturalness.

Practical tip: **higher PSNR or SSIM does not always look better**. Diffusion models often score lower PSNR but better LPIPS and MUSIQ — they look better to a human eye even if individual pixels diverge from the source.

Chapter 12 — Hardware Requirements

| --- | --- | --- | --- |

CPU fallback exists but a single SUPIR image can take tens of minutes. **Apple Silicon punches above its weight thanks to unified memory** — an M3 Max with 64GB runs SUPIR comfortably.

Chapter 13 — Commercial vs Open Source — How to Choose

| Axis | Commercial (Topaz/Magnific) | Open Source (SUPIR/Real-ESRGAN) |

| --- | --- | --- |

| UX | Clean and instant | ComfyUI learning curve |

| Quality ceiling | Good (Topaz conservative, Magnific creative) | SUPIR tops the leaderboard |

| Cost | One-time ~200 dollars or monthly subscription | Free (GPU cost only) |

| Control | A few sliders | Every parameter |

| Batch | Supported within license | Unlimited |

| Updates | Automatic | You manage them |

Recommendations:

- **Casual family photo restoration** -> Topaz Photo AI or Pixelmator Pro

- **Photographer or designer daily workflow** -> Topaz Photo AI + Magnific

- **Research, experiments, total control** -> ComfyUI + SUPIR

- **Bulk automated processing** -> Real-ESRGAN ncnn scripts

- **Manga and illustration** -> Real-CUGAN or Real-ESRGAN Anime

- **Face-centric restoration** -> CodeFormer or GFPGAN

Chapter 14 — Region-Specific Services in Korea and Japan

Korea

- **KakaoTalk Photo Enhance** — the "Sharpen" inside KakaoTalk, light mobile use

- **NAVER Photo Editor** — Naver Clova-based touch-up tools

- **Olive Studio** — AI restoration service from a photo studio chain

- **Camera apps SNOW, Soda, B612** — face beautification first, but also include resolution boosts

Japan

- **Fujifilm X-Ray AI** — Fujifilm photo restoration service

- **Sony Imaging Edge** — Sony camera RAW workflow plus AI denoising

- **Adobe Photoshop Japan edition** — identical feature set with Japanese UI

- **Camera vendor tools**: Nikon NX Studio and Canon DPP both ship ML denoising

The upside of region-specific services is **local payments, customer support, ID-photo culture features**; the downside is models that may not refresh as often as global tools.

Chapter 15 — Real-World Use Cases

Family Photo Digitization

Scan 500 prints at 600dpi, dust removal via Photoshop Neural Filter, CodeFormer face restoration, Topaz Gigapixel 4x upscale, DeOldify colorization for the black-and-white ones. Two to three minutes per image.

E-commerce Sellers

Storefront product images 1000x1000 grown to 4000x4000 — usable for print catalog and full-screen banner simultaneously. Topaz Photo AI batch processing is the workflow standard.

Real Estate Listings

Property photos arrive at mixed resolutions from different phones — Real-ESRGAN or Topaz normalizes them to a consistent size. Automate alongside exposure and color correction.

Manga and Webtoon Remastering

Early-2000s illustrations (typically 1024x1536) into Real-CUGAN 4x produce 4096x6144 output. Lines stay clean while color regions become smooth.

Video Remastering

Old wedding videos (480i SD) into Topaz Video AI Proteus deliver 1080p or 4K. Use the model temporal-consistency options.

CCTV and Surveillance Restoration

Identifying faces in low-res CCTV — CodeFormer with w=0 prefers fidelity. But **this output is not admissible as legal evidence** — diffusion-invented detail can be fake.

Chapter 16 — Ethical and Legal Considerations

Upscaling **synthesizes** information. It creates detail that was never there. A few cautions:

- **Not legal evidence** — upscaled CCTV or surveillance frames are leads at best, not court-admissible

- **Identity preservation for portraits** — CodeFormer w slider closer to 0 stays truer to the source, but diffusion can always draw a slightly different face

- **Copyright** — training datasets may include works the model "remembers". Check the model card license before commercial use

- **GDPR** — sending portraits to a cloud SaaS in Europe requires data processing consent

- **Digital forgery risk** — the same tools can paint in fake detail. Journalism and reporting should disclose when AI upscaling is used

Chapter 17 — Workflow Recommendations by Case

Case A: One Old Family Photo Quickly

Pixelmator Pro ML Super Resolution, one click, done. Five seconds.

Case B: 100 Old Photos in Batch

ComfyUI workflow (GFPGAN + Real-ESRGAN) or Topaz Photo AI Batch.

Case C: Print-Ready Landscape Photo

Topaz Gigapixel 8 High Fidelity, or SUPIR for invented detail.

Case D: Social Content Creator (Creative Freedom OK)

Magnific or Krea — bump the creativity slider for richer detail.

Case E: Manga Collector

Real-CUGAN or Real-ESRGAN Anime in ComfyUI, batch.

Case F: Video Remaster

Topaz Video AI Proteus or Resolve Studio Super Scale.

Chapter 18 — Frequently Asked Questions

**Q. Is higher PSNR always better?**

A. No. PSNR is pixel-difference, so smooth and blurry outputs can win on PSNR. Always look at LPIPS and MUSIQ together.

**Q. Are diffusion models always better?**

A. SUPIR is unbeatable on natural photographs, but for text-heavy or precision-pixel cases (documents, drawings, charts) regression models like SwinIR or HAT are more stable.

**Q. Are free tools enough?**

A. For 80 percent of cases, yes. Real-ESRGAN + CodeFormer + GFPGAN in ComfyUI covers family photos, manga, and general upscale. Topaz is just smoother UI and consistency.

**Q. Can I run SUPIR on a Mac?**

A. An M3 Max or above with 64GB unified memory can (Diffusers MPS backend). On M2 Pro with 16GB, prefer the lighter StableSR or ResShift.

**Q. How does upscaling relate to inpainting?**

A. Inpainting (Stable Diffusion Inpaint) fills holes; upscaling enlarges. That said, diffusion-based upscalers effectively inpaint every pixel.

Chapter 19 — Future — Trends for 2026 and 2027

- **Video diffusion SR** — temporally consistent diffusion video upscaling (Stable Video Diffusion lineage) goes mainstream

- **iPhone on-device models** — Apple Intelligence may ship diffusion upscaling inside the Photos app

- **NPU acceleration** — Snapdragon X Elite, M4, Intel Lunar Lake host NPU-specific upscale models

- **Low-cost cloud GPU** — Modal, RunPod, Together AI expose SUPIR as an API

- **Copyright-safe training data** — Adobe Firefly-style "safely trained" upscalers are likely to appear

- **3D and 4D upscaling** — high-res output for Gaussian Splat and NeRF results

Chapter 20 — Checklist — A First-Run for Beginners

Try this order:

1. Start free — Pixelmator Pro one-click (Mac) or Bigjpg (web)

2. Hit a limit — Topaz Photo AI trial

3. Need control — Stable Diffusion WebUI (Forge) Extras tab for Real-ESRGAN + GFPGAN

4. Going deep — install ComfyUI and a SUPIR workflow

5. Video work — Topaz Video AI trial

6. 100+ images per month — automate (ComfyUI batch or Topaz license)

Epilogue — Restoring Meaning, Not Just Pixels

AI upscaling is interesting because it is not a simple resolution change. The model **knows** that this is a face, this is a blade of grass, this is text — and synthesizes detail accordingly. It restores meaning, not pixels.

Ethical responsibility follows. A restored photo is not what was — it is what plausibly was. When grandfather eye color in a black-and-white family photo becomes brown after colorization, that is the model best guess.

Even so, when an old wedding picture lands at 4K on your living-room TV and your parents smile — the value is real. The meaning was synthesized; the emotion was not.

The 2026 upscaling toolbox is more powerful and more accessible than ever. Topaz, Magnific, or ComfyUI — find the tool that suits your workflow and restore at least one image. That first image is where it starts.

References

- Topaz Labs Photo AI: https://www.topazlabs.com/photo-ai

- Topaz Gigapixel 8: https://www.topazlabs.com/gigapixel

- Topaz Video AI: https://www.topazlabs.com/topaz-video-ai

- Magnific: https://magnific.ai/

- Krea Upscaler: https://www.krea.ai/

- Adobe Camera RAW Enhance / Super Resolution: https://helpx.adobe.com/camera-raw/using/enhance-details.html

- Luminar Neo: https://skylum.com/luminar

- ON1 Resize AI: https://www.on1.com/products/resize-ai/

- Pixelmator Pro: https://www.pixelmator.com/pro/

- Real-ESRGAN (Xintao Wang): https://github.com/xinntao/Real-ESRGAN

- Real-CUGAN: https://github.com/bilibili/ailab/tree/main/Real-CUGAN

- SwinIR (Liang et al, ICCV 2021): https://github.com/JingyunLiang/SwinIR

- HAT (Chen et al, CVPR 2023): https://github.com/XPixelGroup/HAT

- DAT (Chen et al, ICCV 2023): https://github.com/zhengchen1999/DAT

- DRCT (Hsu et al, CVPRW 2024): https://github.com/ming053l/DRCT

- SUPIR (Yu et al, ECCV 2024): https://github.com/Fanghua-Yu/SUPIR

- StableSR (Wang et al, IJCV 2024): https://github.com/IceClear/StableSR

- ResShift (Yue Z. et al, NeurIPS 2023): https://github.com/zsyOAOA/ResShift

- SeeSR (Wu et al, CVPR 2024): https://github.com/cswry/SeeSR

- DiffBIR (Lin et al, ECCV 2024): https://github.com/XPixelGroup/DiffBIR

- PASD (Yang et al, ECCV 2024): https://github.com/yangxy/PASD

- GFPGAN: https://github.com/TencentARC/GFPGAN

- CodeFormer (Zhou et al, NeurIPS 2022): https://github.com/sczhou/CodeFormer

- RestoreFormer: https://github.com/wzhouxiff/RestoreFormer

- waifu2x: https://github.com/nagadomi/waifu2x

- Anime4K: https://github.com/bloc97/Anime4K

- ComfyUI: https://github.com/comfyanonymous/ComfyUI

- Stable Diffusion WebUI Forge: https://github.com/lllyasviel/stable-diffusion-webui-forge

- Bringing Old Photos Back to Life: https://github.com/microsoft/Bringing-Old-Photos-Back-to-Life

- DeOldify: https://github.com/jantic/DeOldify

- NVIDIA RTX Video Super Resolution: https://www.nvidia.com/en-us/geforce/news/rtx-video-super-resolution/

- LPIPS: https://github.com/richzhang/PerceptualSimilarity

- MUSIQ: https://github.com/google-research/google-research/tree/master/musiq