A Generated Face Dataset: AGFD-20K

Can we use generated human faces to train a model? If the model can generate faces that are realistic enough, how can we use it to generate faces we want? What is the mapping between prompt and vary human faces?

Besides, many times we need to construct a good test dataset to test the robustness of the trained model. But selecting images carefully from searching engines is laborious and also facing the problem of cleaning (due to the low quality of the data like unrelated, low-resolution, unbalanced, etc.). How the diffusion model can help us solve this problem?

What's more, fake faces detection is also a essential topic of AI safety. We can generate target faces directly through deep generative models, but just as important is how do we detect them? For example, our logo is A FAKE FACE!

In the past 3 years, we have seen the explosive growth of the Diffusion Model, it now can generate brilliant pictures according to user's prompts. Here I test the diffusion model in the capacity of generating realistic human faces. And purposes of this project are the following two points:

1. Construct a not bad fake faces dataset;
2. Use these data to do something interesting. I don't know if it helps anyone, but it works for me 🤣🤣🤣

The principles I followed in generating faces are:

• Realistic (use Realistic_Vision_V2.0:1 by SG_161222)
• High-resolution (the resolution of all images are 512*512)
• Vary & Balanced (support vary faces and keep good data distribution)

Here is the generated face data and corresponding descriptions:

Attribute	Specific Features	Male	Female	Special Prompt
Age	Child (0-5)	300	300	1 y.o., 3 y.o.
	Teenager (5-18)	600	600	8 y.o., 15 y.o.
	Young people (18-40)	300	300	25 y.o., 35 y.o.
	Middle aged (40-60)	300	300	45 y.o., 55 y.o.
	Old people (60+)	600	600	60,80,100 y.o., Grandma，Grandpa
Emotion	Smile	600	600	smiling, laughing
	Angry	600	600	Angry, pissed-off face, yelling
Occlusion	Only glasses	600	600	glasses,sunglasses,swimming goggles,skiing goggles,
	Only mask	300	300	(masked:1.2), antigas mask
	Only make up	300	300	highly make up, eyeshadow,heavy black eyeliner, joker, Halloween makeup
	Only hands	300	300	put Hand in front of face,put Hand in front of hair
	Complex	300	300	glasses，(masked:1.2),Halloween makeup，put Hand in front of face,put Hand in front of hair
Illumination	Under water	300	300	under water
	Strong light	300	300	(sun behind:1.2), strong sun shine
	Dark	300	300	in the night, dark light, (very dark scene:1.2)
Large pose	Left & right	2000	2000	side view
	Up	300	300	(looking up:1.3)
Hair	Blond	300	300	blond hair
	Bangs	300	300	(Bangs:1.2)
	Bald	300	300	Bald
Others	Moustache	300	-	moustache,sideburns,goatee,front view
	Open Mouse	600	600	(talking loudly:1.4),smile,neutral
	Close Eyes	600	600	(sleepy,close eyes:1.4)
ALL	Origin	20K	20K	-

All the pictures can be downloaded at

Link1: Baidu Disk [extract code: tqxd].

Link2: terabox

Usage

I use this template to get good generation results:

• Prompt:

RAW photo, a close up portrait photo of [year] y.o [sex], [human race],[special prompt],(high detailed skin:1.2), 8k uhd, dslr, high quality, film grain, Fujifilm XT3.

• Negative Prompt:

(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck.

• Hyper-parameter:

DPM++ 2M Karras with 20 steps

CFG Scale 7

Instruction

A reference containing Portraits and Keywords that you can use with Stable Diffusion.

• 👨‍👩‍👧‍👦 Age

  1 y.o., 3 y.o., boy	1 y.o., 3 y.o., girl

  
  

  8 y.o., 15 y.o., boy	8 y.o., 15 y.o., girl

  
  

  25 y.o., 35 y.o., man	25 y.o., 35 y.o., woman

  
  

  45 y.o., 55 y.o., man	45 y.o., 55 y.o., woman

  
  

  60,80,100 y.o., man, grandpa	60,80,100 y.o., woman, grandma

• 😃 Emotion

  smiling, laughing, man	smiling, laughing, woman

  
  

  Angry, pissed-off face, yelling, man	Angry, pissed-off face, yelling, woman

• 🥸 Occlusion

  glasses,sunglasses,swimming goggles,skiing goggles, man	glasses,sunglasses,swimming goggles,skiing goggles, woman

  
  

  (masked:1.2), antigas mask, man	(masked:1.2), antigas mask, woman

  
  

  highly make up, eyeshadow,heavy black eyeliner, joker, Halloween makeup, man	highly make up, eyeshadow,heavy black eyeliner, joker, Halloween makeup, woman

  
  

  put Hand in front of face, put Hand in front of hair, man	put Hand in front of face, put Hand in front of hair, woman

  
  

  glasses，(masked:1.2), Halloween makeup，put Hand in front of face, put Hand in front of hair, man	glasses，(masked:1.2), Halloween makeup，put Hand in front of face, put Hand in front of hair, woman

• 🔆 Illumination

  under water, man	under water, woman

  
  

  (sun behind:1.2), strong sun shine, man	(sun behind:1.2), strong sun shine, woman

  
  

  in the night, dark light, (very dark scene:1.2), man	in the night, dark light, (very dark scene:1.2), woman

• 🔭 Large pose

  side view, man	side view, woman

  
  

  (looking up:1.3), man	(looking up:1.3), woman

• 🦱 Hair

  blond hair, man	blond hair, woman

  
  

  (Bangs:1.2), man	(Bangs:1.2), woman

  
  

  Bald, man	Bald, woman

• 🧑‍💻 Others

  moustache,sideburns,goatee,front view, man	moustache,sideburns,goatee,front view, woman
  	cannot generate 💔

  
  

  (talking loudly:1.4),smile,neutral, man	(talking loudly:1.4),smile,neutral, woman

  
  

  (sleepy,close eyes:1.4), man	(sleepy,close eyes:1.4), woman

Acknowledgement

Thanks the creator of the model for his brilliant work and also thanks his reference models.

Realistic_Vision_V2.0:1 by SG_161222

Future Work

• Although the realistic ability of this version model to generate pictures is enhanced, the generalization is weakened. As reported in MidJourney-Styles-and-Keywords, the Mid-v5 has a better ability to process finer facial details. I will keep focusing on the generated model and use better model to generated more controllable human faces.

Citation

@repo{2023agfd20k,
    title={A Generated Face Dataset: AGFD-20K},
    author={Zhongqi Wang},
    howpublished = {\url{https://github.com/Robin-WZQ/AGFD-20K}},
    year={2023}
}