# What are the best open source photo restoration tools in 2026?

colorizethis.io · September 9, 2026

> The best open source photo restoration tools in 2026 are GFPGAN and CodeFormer for face restoration, Real-ESRGAN for upscaling, DeOldify for...

The best open source photo restoration tools in 2026 are GFPGAN and CodeFormer for face restoration, Real-ESRGAN for upscaling, DeOldify for colorization, and Scanned Image Processing (SIP) workflows built on top of these models. Each solves a different part of the restoration problem, and no single tool does everything well. If you restore old black-and-white family photos, you will almost certainly need to chain two or three of these together: denoise and upscale first, restore faces second, colorize last. This guide breaks down what each tool actually does, where it fails, and how to build a pipeline that produces results worth printing.

## The Direct Answer: Which Tools to Use

**Also worth reading:** [What is the historical photo restoration color workflow and how does AI colorization fit into it?](https://colorizethis.io/knowledge/what_is_the_historical_photo_restoration_color_workflow_and_how_does_ai_colorization_fit_into_it.php) · [How do modern algorithms perform in an AI photo restoration benchmark comparison?](https://colorizethis.io/knowledge/how_do_modern_algorithms_perform_in_an_ai_photo_restoration_benchmark_comparison.php) · [What is hybrid AI manual photo restoration and how does it work for old family photographs?](https://colorizethis.io/knowledge/what_is_hybrid_ai_manual_photo_restoration_and_how_does_it_work_for_old_family_photographs.php)

For most people restoring scanned family photographs, the winning combination is Real-ESRGAN for upscaling and detail recovery, GFPGAN or CodeFormer for damaged faces, and DeOldify (specifically the Artistic or Stable video models) for adding plausible color. All three are free, run locally on a GPU with at least 6-8 GB of VRAM (or slowly on CPU), and are actively maintained on GitHub. CodeFormer, released by researchers at Nanyang Technological University, tends to produce the most convincing faces on severely damaged photographs because it uses a codebook lookup rather than pure generative inference, which reduces the hallucination problem that plagues other models.

If your photos are only mildly degraded, you may not need the full stack. A single pass through Real-ESRGAN at 2x scale, followed by manual dust removal in GIMP or Krita, is often enough for photographs from the 1970s onward. Reserve the heavy generative tools for pre-1950 photographs with genuine damage: cracks, silvering, water stains, or emulsion loss. Applying aggressive AI restoration to a photo that only needs contrast correction is how you end up with a plastic-looking relative who no longer resembles anyone in your family.

## Why Open Source Beats Cloud Services for Restoration

There are three practical reasons to prefer open source tools over subscription cloud services, and one reason that matters more than the others. The decisive reason is fidelity: cloud enhancers optimize for a pleasing result, not an accurate one. Several reviewers testing 12-15 free AI photo enhancers in 2025 and 2026 noted the same pattern — commercial apps smooth skin into porcelain, invent jewelry, and change facial structure to match a generic 'attractive' template. Open source models like CodeFormer include a fidelity weight parameter (typically set between 0.5 and 0.7) that lets you explicitly trade sharpness for accuracy, something no consumer cloud app exposes.

The second reason is privacy and permanence. Family photographs are often the only surviving record of a person, and uploading them to a service whose terms of service can change, whose company can shut down, or whose models train on user uploads is a genuine risk. The third reason is cost at scale: restoring 300 photos through a cloud service at typical per-image or subscription pricing runs $50-200, while a local pipeline costs nothing beyond electricity. The trade-off is honest: local tools require a technical setup, a decent GPU, and patience. If you have three photos and no computer skills, a paid service is the rational choice. If you have an archive, open source wins on every axis that matters.

## Building a Pipeline: Step by Step

A reliable restoration pipeline follows a fixed order, and the order matters because each stage amplifies artifacts from the previous one. Start with a proper scan: 600 DPI minimum for standard prints, 1200 DPI for small or damaged originals, saved as 16-bit TIFF rather than JPEG. JPEG compression artifacts are the single most common reason AI restoration produces smeared results, because the model spends its capacity reconstructing compression noise instead of detail.

Stage one is cleanup. Use GIMP's heal tool or the open source Scantailor-derived tools to remove dust, scratches, and border artifacts manually. This takes 2-5 minutes per photo and dramatically improves everything downstream. Stage two is upscaling with Real-ESRGAN at 2x, not 4x — the 4x model invents substantially more texture and is where most 'nightmare fuel' results come from. Stage three is face restoration: run GFPGAN or CodeFormer only on detected face regions, with CodeFormer's fidelity weight around 0.6 for family photos. Stage four, if desired, is colorization with DeOldify, always applied after restoration, never before, because colorizing a damaged photo locks the damage into colored artifacts. Stage five is a final manual pass in Krita or GIMP for local corrections. Budget 10-20 minutes per photo for a quality result; anyone promising one-click 4K restoration of a badly damaged 1920s portrait is selling you hallucination.

## Comparison: The Major Open Source Options

| Tool | Primary Job | Strengths | Weaknesses | Hardware Needs |
| --- | --- | --- | --- | --- |
| Real-ESRGAN | Upscaling, detail recovery | Fast, stable, few artifacts at 2x | Invents texture at 4x, no face awareness | 4 GB VRAM or slow CPU |
| GFPGAN | Face restoration | Excellent on mild damage, easy one-line setup | Over-smooths skin, changes identity on heavy damage | 4-6 GB VRAM |
| CodeFormer | Face restoration | Best on severe damage, fidelity weight control | Slower, can look waxy at low fidelity settings | 6-8 GB VRAM |
| DeOldify | Colorization | Most natural palette of open options, three model tiers | Occasionally wrong colors (uniforms, cars), no local control | 6 GB VRAM |
| GIMP + plugins | Manual cleanup, finishing | Full control, no hallucination | Labor intensive, no AI restoration natively | Any computer |
| Bringing-Old-Photos-Back-to-Life (Microsoft) | Scratch and blur removal | Handles scratches and non-face regions well | Older model, dated results on faces | 6 GB VRAM |

The Microsoft Bringing-Old-Photos-Back-to-Life project deserves specific mention because it is the only major open source tool that addresses scratches as a first-class problem rather than expecting you to fix them manually. Its global restoration branch handles non-face regions better than GFPGAN, though its face model has been superseded by CodeFormer. A practical 2026 pipeline often uses Microsoft's scratch removal as stage 1.5, between manual cleanup and upscaling.

## Where These Tools Fail: The Hallucination Problem

Every generative restoration model makes things up. This is not a bug that will be patched out; it is inherent to how diffusion and GAN-based models work, since they reconstruct plausible detail from learned patterns rather than recovering actual lost information. On faces, this means CodeFormer at low fidelity settings will give your great-grandmother different eyes, a different nose bridge, or a different jawline than she actually had. PetaPixel's 2025 review of a commercial AI restoration tool documented the same failure mode in consumer products, calling the results 'grotesque nightmare fuel' when the model was pushed too hard.

The practical mitigation is a fidelity-first workflow: always run the gentlest setting that fixes the visible problem, compare before and after at 100% zoom, and ask someone who knew the person whether the result still looks like them. For historical or documentary photographs — Jacob Riis-style social reform imagery, archival records, anything with evidentiary value — generative restoration is arguably inappropriate at all, because the invented detail becomes indistinguishable from the recorded detail. For those images, stick to non-generative cleanup: dust removal, contrast, and despeckling. The distinction between 'making a photo look nice' and 'altering a historical document' is one the open source community takes more seriously than the consumer app market, and you should take it seriously too.

## Colorization: What DeOldify Gets Right and Wrong

DeOldify remains the reference open source colorizer in 2026, nearly seven years after its initial release, and its NoGAN training approach still produces more natural skin tones than most commercial alternatives. Its known failure modes are specific and worth memorizing: it frequently renders 19th and early 20th century military uniforms in the wrong color (Confederate gray and Union blue are common casualties), it struggles with period-accurate car and fabric colors, and it tends to over-saturate foliage. None of these are fixable automatically, because the model has no way to know what color the original object was.

The correct mental model is that AI colorization produces a plausible starting point, not a finished colorization. Professional colorists working on archival projects spend hours per image correcting DeOldify's output using period documentation — paint chips, fabric samples, written descriptions. For family photos where exact accuracy matters less than emotional resonance, DeOldify's output at 70-80% opacity blended over the original grayscale often looks more natural than its raw output, which tends toward oversaturation. If you want guided colorization with manual control, the open source G'MIC plugin suite for GIMP includes interactive colorization tools that trade automation for precision.

## Cost, Hardware, and When to Act

The software is free; the hardware is not, necessarily. A used GPU with 8 GB of VRAM — an RTX 3060 or equivalent — runs the full pipeline comfortably and can be found for $150-250 in 2026. Running the models on CPU works but takes 3-10 minutes per image instead of 10-30 seconds, which is tolerable for a dozen photos and miserable for an archive of 500. Cloud GPU rental through services like RunPod or Vast.ai costs roughly $0.20-0.50 per hour, meaning a 300-photo archive can be processed for under $5 of compute if you batch it.

When to act is straightforward: scan deterioration is real and physical. Photographic prints from 1890-1950, particularly silver gelatin prints, are actively degrading — emulsion cracking, silver mirroring, and sulfur toning damage compound every year, and color prints from the 1960s-80s fade fastest of all because of unstable chromogenic dyes. Scanning at 600+ DPI costs nothing after the initial equipment purchase, and a scan preserves the current state even if you never run restoration on it. Scan first, restore later. The restoration tools will keep improving; the originals will not.

## Common Mistakes and How to Avoid Them

The most frequent mistake, reported consistently across user forums and reviewer testing, is over-processing: stacking 4x upscale, maximum face enhancement, and full colorization on a photo that needed a contrast adjustment. The result is a hyper-smooth, oversaturated figure that resembles the subject the way a police sketch resembles a suspect. Work at the lowest intervention level that solves your actual problem, and stop there.

The second mistake is scanning wrong: 300 DPI JPEGs from a flatbed's default settings throw away information you cannot recover later. The third is colorizing before restoring, which bakes damage into color. The fourth is trusting face restoration blindly on people you never met — CodeFormer will confidently produce a plausible face, and if you have no reference, you cannot tell how much is invention. Cross-check against other photographs of the same person whenever possible. The fifth mistake is discarding the original: always keep the untouched scan as your master file, export restoration results as separate copies, and name them systematically. In five years, when restoration models improve again, you will want to rerun the pipeline from the clean source, not from a 2026-era AI output that already contains hallucinated detail.

## Quick answers

### Is GFPGAN or CodeFormer better for restoring old family photos?

CodeFormer is better for severely damaged faces because its fidelity weight parameter lets you control the accuracy-versus-sharpness tradeoff, and its codebook approach reduces hallucination. GFPGAN is faster and easier to set up but tends to over-smooth skin and can alter facial identity on heavy damage. Many pipelines use GFPGAN for mild cases and CodeFormer for bad ones.

### Can I run these open source restoration tools without a GPU?

Yes, all the major models run on CPU, but expect 3-10 minutes per image instead of seconds. For a handful of photos this is fine; for an archive of hundreds, a used GPU with 8 GB of VRAM (around $150-250) or cloud GPU rental at roughly $0.20-0.50 per hour is worth it.

### Does AI colorization know the real colors of old photos?

No. DeOldify and similar models predict statistically plausible colors based on learned patterns, not recovered information. Uniforms, cars, and fabrics are common failure cases. Treat the output as a starting point and correct it manually if historical accuracy matters.

### What resolution should I scan old photos at before restoration?

Scan at 600 DPI minimum for standard prints and 1200 DPI for small or damaged originals, saved as 16-bit TIFF. Avoid JPEG, since compression artifacts degrade AI restoration results and cannot be removed downstream.

### Are open source restoration tools better than paid apps like commercial enhancers?

For accuracy and control, yes — open source tools expose fidelity parameters and run locally, while consumer apps optimize for pleasing results and often invent details or smooth faces unnaturally. For convenience with just a few photos and no technical skills, a paid cloud service is the simpler choice.

Canonical: https://colorizethis.io/knowledge/what_are_the_best_open_source_photo_restoration_tools_in_2026.php
Markdown: https://colorizethis.io/knowledge/what_are_the_best_open_source_photo_restoration_tools_in_2026.php/index.md