Best AI Photo Restoration Tools in 2026

AI photo restoration takes a damaged, faded, or low-resolution photograph and reconstructs missing detail, removes physical damage like scratches and creases, sharpens blurred faces, and optionally adds color to black-and-white images. The input is a scanned or photographed image showing one or more types of degradation. How it works: These models analyze degraded pixels, infers what the original content looked like using patterns learned from millions of clean photographs, and outputs a repaired version.

With AI photo restoration tools you can restore family photos from the 1940s or build consumer apps that charge per restored image, you can set yo an archival digitization projects processing thousands of damaged prints.

however, Nearly every tool listed below can be self-hosted on a rented GPU for pennies per image,however aifreeforever photo restore uses GFPGAN and can be used completely free.

Top AI Photo Restoration Tools

1. GFPGAN

GFPGAN (Generative Facial Prior GAN) is the most widely used face restoration model. Developed by Tencent ARC Lab and published at CVPR 2021, GFPGAN takes a degraded face image and reconstructs sharp facial features by leveraging generative priors from a pretrained StyleGAN2 face generator.

• The model detects and aligns faces automatically, restores fine-grained detail in eyes, teeth, and hair, and outputs a 2x-upscaled result.
• GFPGAN handles compression artifacts, motion blur, Gaussian noise, and age-related photo deterioration without requiring any information about the specific type of degradation. That blind restoration capability is what makes the model practical for real-world use, where you rarely know exactly what happened to a photograph.
• The limitation is scope. GFPGAN operates on faces only. Background regions, clothing, and non-facial elements pass through without meaningful restoration. For a full-image pipeline, you typically pair GFPGAN with a general-purpose upscaler like Real-ESRGAN for the background, then composite the restored face back into the enhanced frame.

You can restore old photos using GFPGAN at zero cost by self-hosting on any NVIDIA GPU with 4 GB+ VRAM. Inference takes roughly 1–2 seconds per image on a consumer RTX card.

2. CodeFormer

CodeFormer is the second major face restoration model from the Tencent research ecosystem, published at NeurIPS 2022. Where GFPGAN uses GAN-based generative priors, CodeFormer takes a different architectural approach: a codebook lookup transformer. The model learns a discrete codebook of high-quality face components during training, then matches degraded input regions to the closest clean codebook entries to reconstruct the face.

The key feature is a tunable fidelity parameter (w) that controls the balance between restoration quality and faithfulness to the original. At w=0, CodeFormer generates the sharpest possible output but may alter facial identity. At w=1, the model stays closer to the input pixels, producing a more conservative restoration. In practice, values between 0.5 and 0.7 produce the most useful results for archival work where preserving the person’s recognizable features matters.

CodeFormer handles the same degradation types as GFPGAN and generally produces smoother, more natural skin textures, while GFPGAN tends to recover sharper eye and teeth detail.

3. SUPIR

SUPIR (Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild) is the heaviest restoration model in this list and the only one that operates effectively on both faces and full scenes. Developed by the XPixel Group, SUPIR is built on top of the SDXL diffusion architecture and was trained on 20 million high-resolution images, each paired with text annotations.

Multiple deployable variants exist: SUPIR-v0Q (higher perceptual quality, more hallucination) and SUPIR-v0F (higher fidelity, closer to input). Choose the variant based on whether you prioritize visual impact or source accuracy.

4. Bringing Old Photos Back to Life

Bringing Old Photos Back to Life is the only model in this list purpose-built for physical damage. Developed by Microsoft Research Asia and published at CVPR 2020 as an oral paper, the model uses a four-stage pipeline: global restoration through paired VAEs, face detection and bounding, face-specific enhancement using a dedicated network, and a final blending step that composites the improved faces back into the restored image. For large archival scans, you get better results by running the restoration at native resolution and then passing the output through a dedicated upscaler.

5. DDColor

DDColor handles automatic colorization of black-and-white photographs. Published at ICCV 2023, the model uses a dual-decoder architecture built on a ConvNeXt backbone to optimize learnable color tokens, producing photo-realistic color for historical grayscale images.

DDColor ships in four pretrained variants:

• ddcolor_modelscope — the default general-purpose model, trained on broad image data
• ddcolor_paper — the original research checkpoint, slightly different color profile
• ddcolor_artistic — trained with extended artistic image data, produces more saturated and stylized colors
• ddcolor_paper_tiny — a lightweight version for faster inference on resource-constrained hardware

You can colorize photos for free by running DDColor locally or through hosted APIs.

6. SwinIR

SwinIR applies the Swin Transformer architecture to image restoration tasks including super-resolution, JPEG artifact removal, and general denoising. Published as a research model by Jingyun Liang, SwinIR processes the full image (not just faces) and excels at recovering fine texture detail like fabric weave, wood grain, and foliage in landscape photographs.

SwinIR operates at fixed upscaling factors (2x, 3x, 4x) and does not perform face-specific restoration. For old photos with both face damage and general degradation, the practical workflow is to run a face-specific model first (GFPGAN or CodeFormer), then apply SwinIR to the entire output for background detail recovery and artifact cleanup.

7. NAFNet

NAFNet (Nonlinear Activation Free Network) takes a minimalist approach to image restoration. Developed by Megvii Research, NAFNet removes nonlinear activation functions entirely from the network and replaces them with simple channel attention and element-wise multiplication. The result is a model that achieves competitive denoising and deblurring performance with a smaller parameter count and faster inference than attention-heavy alternatives.

NAFNet is useful for photographs where the primary degradation is camera shake blur or high-ISO sensor noise rather than physical damage. Scanned photos that are clean but soft, or digital photos shot in low light, respond well to NAFNet’s denoising pipeline. The model does not handle scratches, tears, or colorization.

8. DeOldify

DeOldify is a GAN-based colorization model that has been in active use since 2018, making it one of the oldest tools in this roundup. The image variant takes grayscale input and produces a colorized output using a self-attention GAN architecture.

DeOldify tends to produce more muted, historically plausible colors compared to DDColor’s more vivid palette. For archival projects where you want the colorization to look period-appropriate rather than hyperrealistic, DeOldify is often the better choice. The model also has a dedicated video colorization variant for footage restoration.

9. VQFR

VQFR (Vector-Quantized Face Restoration) is Tencent ARC’s codebook-based face restoration model, a precursor to CodeFormer’s approach. VQFR uses a vector-quantized dictionary to store high-quality face components and reconstructs degraded inputs by looking up the nearest clean matches.

VQFR produces sharper output than GFPGAN on heavily degraded inputs where almost no facial structure remains visible, but the model is more aggressive in its reconstructions and can alter facial identity more noticeably. For photos where the face is severely damaged and the priority is producing any recognizable result rather than a perfect likeness, VQFR fills a gap that the more conservative models do not.

10. Topaz Labs Dust & Scratch + Colorization

Topaz Labs offers the only paid restoration tools in this list. The dust-and-scratch-v2 model detects and removes physical print damage (dust specks, film scratches, water stains) from scanned photographs. The image-colorization model adds color to grayscale inputs using Topaz’s proprietary processing pipeline.

Topaz moved to subscription-only pricing in October 2025. Current plans:

Additional open-source restoration models

Beyond the ten primary entries, 15 additional open-source models cover specialized restoration tasks:

• GPEN — blind face restoration using GAN priors with a different architecture than GFPGAN. Produces slightly softer output, useful as an alternative when GFPGAN over-sharpens on certain face types.
• BigColor — colorization using a generative color prior, producing broader color palettes than DDColor on some landscape scenes.
• MAXIM — Google Research’s multi-axis MLP for denoising, dehazing, deblurring, and enhancement. Handles multiple degradation types in a single model.
• IFAN Defocus Deblur — specialized model for removing defocus (out-of-focus) blur. Effective on scanned photos where the original camera missed focus.
• Color Matcher — corrects white balance and matches color palettes between images. Useful as a post-processing step after colorization to ensure consistency across a batch.
• SUPIR-v0Q and SUPIR-v0F — variant deployments of the SUPIR model optimized for quality (v0Q) and fidelity (v0F) respectively.
• Night Enhancement — enhances photos taken in low-light conditions, recovering shadow detail and reducing noise.
• PASD Magnify — pixel-aware stable diffusion super-resolution. Academic-stage model, useful for extreme upscaling with creative detail generation.
• InstructIR — an instruction-following restoration model. You describe the degradation in natural language (“remove the blur and increase sharpness”), and the model applies the corresponding restoration.
• BSRGAN — trained on a broader degradation model than Real-ESRGAN, making it more robust to unusual combinations of blur, noise, and compression in real-world photos.
• Object Remover — inpainting model for removing unwanted objects from photos, useful for cleaning up scans where a sticker, tape, or other obstruction covers part of the image.
• Magic Style Transfer — restyles an image based on another reference image, less restoration-focused but included in restoration collections for creative re-imagining of old photos.
• Restore Image — a Flux Kontext-based restoration and colorization model combining multiple restoration tasks in a single call.
• Real-ESRGAN — technically an upscaler rather than a restoration model, but so commonly paired with GFPGAN in restoration pipelines that it belongs in any restoration toolkit. Real-ESRGAN handles background and non-facial upscaling while GFPGAN restores the faces.
• Clarity AI Upscaler — an open-source upscaler positioned as a free alternative to Magnific. The base clarity-upscaler model is free to self-host, with a paid app and API (~$0.02/run) for managed access. Subscriptions run $29–$299/month.

A lot of models listed in this article are available as open-source weights you can download and run locally, with the sole exception of Topaz Labs’ proprietary models.