How to Use Replicate CodeFormer: Turn Blurry Photos into Crystal-Clear Portraits

Those old family photos sitting in dusty albums—you know, the ones where faces are so blurry you can barely make out who's who? I've been there, and I've been testing AI photo enhancement tools to see which ones actually deliver on their promises.

CodeFormer caught my attention after I put it through its paces alongside a bunch of other face restoration tools. What sets it apart is how well it handles both vintage family photos and AI-generated faces that need cleaning up. The numbers back this up too—CodeFormer has been adopted across more than 20 online platforms and API services, racking up 46.5 million runs from users who clearly found it useful.

The pricing is what really makes this tool accessible. Each run costs approximately $0.0027 on Replicate, which means you can restore about 370 photos for just a dollar. Processing typically wraps up in 3 seconds[-3], so you're not sitting around waiting for results.

I'll walk you through exactly how to set up and use CodeFormer to rescue those blurry images. Whether you're trying to salvage old family memories or clean up photos for professional work, this guide covers what you need to know.

Set Up CodeFormer on Your Machine

Running CodeFormer locally gives you complete control over your photo restoration process, plus you won't need internet access once everything's installed. The setup involves a few technical steps, but I'll walk you through each one.

Your system needs Pytorch ≥ 1.7.1 and CUDA ≥ 10.1 before you start. Once that's sorted, here's what you need to do:

1. Clone the repository:
2. git clone https://github.com/sczhou/CodeFormer
cd CodeFormer

3. Create and activate a dedicated environment:
4. conda create -n codeformer python=3.8 -y
conda activate codeformer

5. Install dependencies:
6. pip3 install -r requirements.txt
python basicsr/setup.py develop

7. Download the essential pretrained models:
8. python scripts/download_pretrained_models.py facelib
python scripts/download_pretrained_models.py CodeFormer


The commands above will automatically place the models where they need to go. If you want better face detection capabilities, you can also add dlib:

conda install -c conda-forge dlib


The whole installation process usually takes 15-20 minutes, depending on your internet speed and how fast your system is. Once it's done, you're ready to start turning those blurry portraits into sharp, clear images.

Run CodeFormer for Different Use Cases

With CodeFormer installed, you can tackle different types of photo restoration depending on what you're working with. I've found the tool handles various scenarios surprisingly well.

For cropped and aligned face images (the standard 512x512 size), this command works best:

python inference_codeformer.py -w 0.5 --has_aligned --input_path [image folder]


When you're dealing with full images rather than just facial crops, I use this setup:

python inference_codeformer.py -w 0.7 --input_path [image folder]


You can push the quality further by adding --bg_upsampler realesrgan to enhance backgrounds, or throw in --face_upsample to sharpen facial details even more.

The -w parameter (fidelity weight) is where things get interesting. It ranges from 0 to 1, and honestly, getting this right makes all the difference. Lower values like 0.5 give you better detail enhancement, while higher values around 0.7 or 1.0 stick closer to the original image. I usually start with 0.5 and adjust from there.

For video files, you'll need ffmpeg first:

conda install -c conda-forge ffmpeg


Then process your mp4, mov, or avi files with:

python inference_codeformer.py --bg_upsampler realesrgan --face_upsample -w 1.0 --input_path [video path]


CodeFormer handles noise, compression artifacts, and blurriness pretty effectively. Processing runs 2-5 seconds per image, which makes batch jobs totally manageable. (Just don't expect miracles if your source material is extremely degraded—even AI has its limits.)

Explore Advanced Features and Online Options

CodeFormer goes beyond basic face restoration with a couple of specialized features that are genuinely useful for specific scenarios.

Face inpainting is handy when you're dealing with photos that have actual damage—scratches, tears, or missing chunks. You mark the damaged areas with a white brush, and CodeFormer fills them in:

python inference_inpainting.py --input_path [image folder]


Face colorization brings old black and white photos to life, which honestly impressed me more than I expected:

python inference_colorization.py --input_path [image folder]


While local installation gives you complete control, sometimes you just want to test a few images without the setup hassle. Here are your online options:

• Replicate - The most reliable option I've used, with consistent 3-second processing and the same pricing we covered earlier
• Hugging Face Space - Free demo access, though it can be slower during peak times
• Stable Diffusion WebUI - Available as an integration if you're already using that ecosystem

The Replicate Codex is worth checking out if you want to compare CodeFormer against similar models. It's a searchable database where you can filter by price, creator, or specific capabilities—useful for finding alternatives or complementary tools.

One approach I've seen work well is combining CodeFormer with other restoration tools. Some users run images through both CodeFormer and GFPGAN in sequence. It takes longer, but the results can be more polished than using either tool alone.

The weight parameter deserves another mention here: use 0 for fixing severe distortion and 1 for adding crispness without geometric changes. I usually start at 0.7 and adjust from there based on what the image needs.

Conclusion

After putting CodeFormer through various restoration scenarios, I'm impressed with how well it handles the job. It's one of those tools that actually delivers on what it promises—turning genuinely unusable photos into something you'd be happy to share.

The tool strikes a nice balance between being powerful enough for serious restoration work and accessible enough that you don't need a computer science degree to use it. Whether you go with the local setup for full control or stick with online platforms for convenience, the results are consistently solid.

What I appreciate most is the flexibility. The weight parameter gives you real control over how aggressive the restoration gets, which matters when you're working with family photos that have sentimental value. Sometimes you want maximum enhancement, other times you just need to clean things up without changing the character of the original image.

I've tested plenty of photo restoration tools over the past year, and CodeFormer has earned its spot in my regular toolkit. It's particularly good for batch processing—I recently used it to clean up about 50 old family photos in one go, and the consistency was impressive.

The key is experimenting with those weight settings until you find what works for your specific photos. Lower values around 0.5 work great when you want maximum quality improvement, while higher values preserve more of the original character. Once you get the hang of it, you'll have those dusty album photos looking sharp again.

Key Takeaways

Transform your blurry, damaged photos into crystal-clear portraits using CodeFormer, an AI-powered face restoration tool that's both affordable and highly effective.

• Ultra-affordable processing: At just $0.0027 per run on Replicate, you can restore about 370 photos for only $1 with 3-second processing speed.

• Flexible installation options: Choose between local setup for maximum control or convenient online platforms like Replicate and Hugging Face Space.

• Master the fidelity weight parameter: Use lower values (0.5) for higher quality enhancement or higher values (0.7-1.0) to preserve original image characteristics.

• Beyond basic restoration: Access advanced features like face inpainting for damaged areas and colorization to bring black-and-white photos to life.

• Versatile media support: Process individual images, batch folders, or entire videos with background upsampling for comprehensive enhancement.

CodeFormer's proven track record of 46.5 million runs across 20+ platforms demonstrates its reliability for both professional photographers and personal photo restoration projects.

FAQs

Q1. How effective is CodeFormer in restoring blurry images? CodeFormer is highly effective in restoring blurry images. It uses advanced AI algorithms to enhance image quality, sharpen details, and reduce blurriness while maintaining a balance between improvement and preservation of the original image characteristics.

Q2. What is the cost of using CodeFormer for photo restoration? CodeFormer is very affordable. On platforms like Replicate, it costs approximately $0.0027 per run, which means you can process about 370 photos for just $1. This makes it accessible for both personal and professional use.

Q3. Can CodeFormer be used for enhancing video quality? Yes, CodeFormer can be used to enhance video quality. After installing ffmpeg, you can process video files (mp4, mov, or avi) using a specific command that includes background upsampling and face enhancement features.

Q4. What is the fidelity weight parameter in CodeFormer, and how does it work? The fidelity weight parameter in CodeFormer, ranging from 0 to 1, controls the balance between enhancement quality and original image preservation. Lower values (around 0.5) produce higher-quality outcomes with better details, while higher values (0.7-1.0) maintain greater fidelity to the original image.

Q5. Are there any online options available for using CodeFormer? Yes, there are several online options for using CodeFormer. These include Replicate, which processes images in about 3 seconds per prediction, Hugging Face Space, which offers a free online demo, and integration with Stable Diffusion WebUI. These platforms make CodeFormer easily accessible without local installation.

‍