How To Fix Eyes In Stable Diffusion [3 Working Methods]

Are you struggling to generate beautiful eyes in Stable Diffusion? 

This is a common problem many users face as sometimes eyes aren’t generated properly and you get bad eyes that look weird or straight-out horrific. 

But luckily you can fix eyes in Stable Diffusion through various methods that I’ll be sharing in this guide. 

Whether you want to fix eyes on an image you already generated or want to never face the issue of bad eyes, I’ve covered everything for you. 

So, let’s get started. 

How To Fix Eyes In Stable Diffusion 

While there are plenty of methods to fix eyes in Stable Diffusion, I’ll be only covering the ones that work in most cases and aren’t very time-consuming. 

Here are the different methods you can follow to fix bad eyes in Stable Diffusion: 

Use Inpainting [Recommended]

My recommended method to fix eyes in Stable Diffusion is using the inpainting feature to mask the eyes and then write a prompt to fix them. 

This is by far the quickest and easiest method to fix eyes in Stable Diffusion. This is a tried-and-tested method and I’ve found myself using it to often fix eyes that look bad. 

The concept is very simple: 

You upload an image you generated to the inpainting tab in img2img and draw a mask over the eyes. Then, you can enter a prompt to fix the bad eyes. 

Here’s an image with bad eyes that we’ll fix as a demonstration of this method: 

In Automatic1111, click on the Send to Extras button right below the generated image. 

If you’re uploading an old image, go to the img2img tab and click on the inpaint tab and directly upload your image there. 

Now that our image is in the inpaint tab, draw a mask over the eyes and cover them completely. 

All that’s left now is to write a positive and negative prompt to fix the bad eyes. Here’s the prompt I’ve used in this example: 

Positive Prompt: 

beautiful hazel eyes, highly detailed eyes, realistic eyes, thin eyebrowsCopy

Negative Prompt: 

ugly eyes, imperfect eyes, deformed pupils, deformed iris,Copy

You’ll notice that my prompts are very simple and short. The reason behind this is that writing a long or complicated prompt for fixing eyes would make things even worse. 

Since we’re not generating an entire image and are only fixing bad eyes, you can just enter a short prompt and get good results. 

Here are the configuration settings I’ve used for Inpainting:

• Sampler: DPM++ 2M SDE Karras
• Mask Mode: Inpaint Masked
• Masked Content: Original
• Inpaint Area: Only Masked

I haven’t changed the other settings and left them as is. However, you can experiment with the Sampler and test which one gives a better output. 

For the model, you can use the same one you used to generate the image. In this case, I’m using epicRealism checkpoint model which I’ve also used to generate the image. 

Here are the results after Inpainting: 

The results are pretty impressive and now the eyes look a lot better than the original image. 

You can further improve the image by using a negative embedding in your prompt while inpainting which brings us to our next method. 

Related: How To Change Clothes In Stable Diffusion

Use Embeddings & LoRA Models

Using Embeddings or LoRA models is another great way to fix eyes in Stable Diffusion. The beauty of using these models is that you can either use them during image generation or use them during inpainting to fix a badly generated eye. 

There are plenty of Negative Embedding (or Textual Inversion) models that will help you avoid bad eyes in your images. 

Here are some of my recommended Embeddings for fixing bad eyes: 

• EasyNegative
• Fast Negative Embedding

These embeddings are trained on a wide variety of unwanted aspects and artifacts and will not only help you fix bad eyes but also fix hands, legs, mouth, and more. 

Here’s the result of fixing eyes using Embeddings in inpainting:

Besides Embeddings, there are plenty of LoRA models you can use to generate beautiful and highly detailed eyes. 

Here are some of my recommended LoRA models to generate beautiful eyes in your images: 

• Eyes / (male) from Polyhedron
• Eyes / (female) from Polyhedron
• Perfect Eyes XL

Apart from this, there are also LoRA models like Realistic Eyes that are specifically made for inpainting. So if you want to inpaint bad eyes, you can use this model. 

Related: How To Generate Consistent Faces In Stable Diffusion

Use Proper Prompts

The last method to fix eyes in Stable Diffusion is using proper prompts to avoid bad eyes. 

While prompting doesn’t always guarantee you’ll never face issues with eyes, you can still reduce such occurrences with the use of the right prompts. 

Entering words either in the positive or negative prompt can be helpful in generating beautiful eyes. 

I like adding certain words in my positive prompts which is helpful in generating good eyes. 

Here are some words you can enter in your positive prompt to generate better eyes in Stable Diffusion: 

• Realistic eyes
• Beautiful eyes
• Highly-detailed eyes
• Perfect eyes
• Hazel eye
• Photorealistic eyes
• Brown eyes

However, if you want to include words in your negative prompt, you can include these:

• Deformed iris
• Deformed pupils
• Morbid
• Oval iris
• Collapsed eyeshadow
• Multiple eyebrows
• Skewed eyes
• Imperfect eyes
• Ugly eyes

These words will help you fix a lot of issues you’d usually face with bad eyes. 

Overall, a combination of the right positive and negative prompts along with a negative embedding goes a long way in generating beautiful eyes in Stable Diffusion. 

Related: How To Swap Faces In Stable Diffusion

Conclusion

Fixing bad eyes in Stable Diffusion isn’t difficult and can be done in various ways. In this guide, we explored three methods that work well in fixing eyes and ensure you never generate bad eyes in the first place. 

Apart from the methods I’ve shared, you can use ControlNet to fix eyes but I find that process to be time-consuming considering the diminishing returns. 

You can find videos of it on YouTube but the results are pretty much the same as you’d get with Inpainting it takes half the time. 

If you have any questions regarding the methods shared in this guide, feel free to ask them in the comments below.