In just a couple of years, AI tools such as Stable Diffusion, Midjourney, Dalle-2, and others have come a long way in making realistic human looking faces. Unlike the mutant hands we usually get, faces are becoming near impossible to tell the real vs the AI fakes these days.
Today, we’ll be exploring which Stable Diffusion checkpoint model gets the best results for creating both men’s and women’s faces. Using a simple prompt and the same settings, we’ll try out different models and compare how each fares to make some basic faces.
The Most Beautiful Women & Men According to AI
So here’s the criteria I’ll be looking at to judge each model for how well they produce the most beautiful AI women and men’s faces.
- Realism – Let’s be honest, if it looks like a cartoon it’s a fail.
- Aesthetics – We’re looking to make AI models, not uglies.
- Variety – If a model keeps generating the same basic face, people will quickly begin to identify it as an AI face, not a real human face.
- Technical Accuracy – Are the eyes the same size, teeth straight, etc.?
So, let’s begin
Creating Beautiful Faces in Stable Diffusion – The Setup
For this simple test, I’m going to use Easy Diffusion. I know Automatic1111 has some advanced features that haven’t yet been implemented into Easy Diffusion, but I don’t really need any of those for this test.
Here’s the Checkpoint models I’ll be comparing:
- absolutereality_v16
- aZovyaPhotoreal_v2
- cyberrealistic_v32
- majicmixRealistic_v6
I chose these four because they’ve all shown themselves capable of making photorealistic human faces that could fool the average person. So let’s begin.
Here’s the configuration I set:
For each image I used a random seed. I chose to use the DPM++ 2M (Karras) Sampler because I’ve found that one to get consistantly good results on nearly any model. I’m sure that certain checkpoints work better with other samplers, but to give it a fair shakedown, I had to pick one that they all perform well with.
The only LoRA I chose to use is the add_detail LoRA which I set to 0.7. I really like how this LoRA improves the level of detail of skin, hair, eyes, etc when using it to generate people. Actually, it’s a LoRA I use for pretty much everything these days.
I didn’t use any VAE or FaceFix models, since this test is to see how well the models do on their own without another layer trying to fix the faces.
I kept the image size set to the default of training at 512×512 pixels and the guidance scale at 10.7. Also the number of steps I kept at 23 for all models. I also used the negative embeddings BadDream and UnrealisticDream.
For each model I ran the test three times without clip skip, and three times with it turned on. Here were the results:
The Best Stable Diffusion Model to Make Beautiful Women’s Faces
So here’s the prompt I used to generate these faces: masterpiece, (((most beautiful woman))), breathtaking, beauty, gorgeous, smile, Ultra-HD, 4K. Now, for each model’s results.
absolutereality_v16 Model
Without using clip skip:
With clip skip enabled (notice the teeth in the smiles now):
AzovyaPhotoreal_v2
Without using clip skip
Now with clip skip (better variety of faces using clip skip)
majicmixRealistic_v6
Without using clip skip (obviously trained on mainly Asian faces)
With clip skip enabled (seems even more cartoonish)
cyberrealistic_v32
With no clip skip
With clip skip (better variety)
And now for the men…
The Best Looking Men’s Faces According to AI and Stable Diffusion
For the men’s faces, I’m going to use a very similar prompt to what I did for the women’s faces. However I added in some masculine words like “manly” & “beard” because I know that SD tends to create very effeminate men’s faces if you don’t. I think the issue is simply too much training data of women’s faces compare to men’s.
Here were the results.
absolutereality_v16 Model
Without clip skip
With clip skip (noticeably more variety and actual smiles)
AzovyaPhotoreal_v2
Without clip skip (looks like the exact same guy three times)
With clip skip (noticeably better variety and again, teeth in teh smiles)
majicmixRealistic_v6
Without clip skip (noticeably more Asian looking than the other models, slightly cartoony)
With clip skip (even more cartoonish, and quite feminine features)
cyberrealistic_v32
Without clip skip (pretty similar-looking faces in all three).
With clip skip (a bit better variety, especially of hair styles)
Conclusion
All four stabile diffusion checkpoint models are able to generate realistic faces. Of the four, I would say that majicmixRealistic is the loser. Compared to the other three, the faces definitely looked more like a cartoon. If you’re going for a sort of cyber semi-realistic anime look, this might be just what you are going for, but it’s not as likely to fool anyone into thinking it’s a real photograph.
aZovyaPhotoreal was able to create highly-realistic faces that are quite good looking I must say. However, it does lack variety in how the faces look while using a random seed. I can certainly be used for this, but it’s not a model you can expect to get a great variety from a single prompt without changing up the wording each time. Otherwise they tend to look pretty cookie-cutter. But, it’s only version 2, so we’ll see what future training might do to add variety into the mix.
The two winners are absolutereality_v16 and cyberrealistic_v32. It’s honestly hard to choose between the two Each has their own style and look, but both produce beautiful AI faces that are realistic and have a fair bit of variety out of a single prompt. This didn’t come as a huge surprise as they’re version 16 and 32 respectively, and obviously have undergone extensive training far beyond the others.
The Winner for Most Beautiful AI Woman / Men’s Faces
If I had to choose just one, I’d probably go with cyberrealistic_v32 for making beautiful AI faces. That’s not just based on this test, but also in my day-to-day use, I’ve found that when I need to make real-looking people, it’s the checkpoint that constantly delivers results.
Have another model you think works better? Let me know in the comments below and I’ll include it in my next testing.