Over the past several months, I have dedicated myself to perfecting a revolutionary AI-driven solution – the FutGAN. In my relentless pursuit, I encountered numerous challenges while crafting a new-generation facepack for FM. While initial attempts yielded some successes, some persistent flaws and issues demanded a comprehensive solution.
Today, I am thrilled to unveil the FutGAN – a cutting-edge Stable Diffusion model meticulously trained on FIFA player images. This breakthrough tool is a game-changer for managers seeking unparalleled realism and diversity in their gaming experience. Through trial and error and previous attempts that resulted in unsatisfactory or labor-intensive outcomes, FutGAN promises to redefine how we approach newgen face creation.
What sets FutGAN apart is its remarkable ability to deliver realistic and unique faces, ensuring that newgens no longer appear repetitive or overly cartoony. Thanks to its versatile prompt-based generation, managers can now customize every aspect of their newgen players. Whether it's long or short hair, clean-shaven faces, full beards, or even neck tattoos – FutGAN empowers you to essentially give life to your newgen-infested squads.
The link to the model can be seen here: https://civitai.com/models/245295
How do I create FutGAN faces?
Generating these images may seem complicated at first, but here is a guide to help you start:
There will be a Create button on the right side of the model. This will be the first step of your creation process.
A new page on the left will now pop up, with details on how to generate your data.
For this LoRA model, I recommend using epiCPhotoGasm. This checkpoint model provides realistic images with only short prompts, unlike other models that need longer prompts. However, you can use other models such as RealisticVision or epicRealism, you probably need to modify other details such as CFG and steps (will be provided for epiCPhotoGasm for this guide).
The recommended weight for this LoRA is 0.85-0.95; anything lower will generate images from epiCPhotoGasm (too generic), and anything higher will provide deep-fried images.
Luckily, we don't have to provide long prompts for epiCPhotoGasm, so a prompt including age, nationality/ethnicity, hairstyle, facial hair, and other characteristics can be easily added. To produce consistent results, add "plain white background" at the end of the prompt.
Note: if you want to generate faces for players ages 19 or below, it is advised to use 20yo instead; new community guidelines have been implemented that prohibit the use of certain keywords.
You don't need to do plenty of work for the negative prompt. Just copy this text instead:
white hair, turban, cap, headwear, hat, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime), text, 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, UnrealisticDream
Generally, it is not necessary to tinker with these settings, but I will give a brief description of these and recommended values.
For this model, a square aspect ratio should be used.
A CFG scale is a value that determines how much of the prompt will be followed; a lower value gives you a more creative and loose image and a higher value gives you more precise and generic (other times, deep-fried) images. I recommend using values between 4.0 and 7.0.
DPM++ 2M Karras is the best sampler for this model.
I don't see any evident changes when I set the step number, so any value is fine. The general rule is lower values give you less-effort images and higher values take too much computing power and time just for one image, I'd recommend values between 15-30.
I would set a random seed for each generation just in case I'm not satisfied with the produced image. If you want to retain a consistent face and change facial characteristics for a specific player, you can input the UID of that player as a reference.
VAE is sometimes baked into realistic models, so you probably don't have to add that.
You can choose to have 1 image to 10 images, and there you have it! FutGAN faces.
Other tools that can be useful to you:
Background Remover: https://www.remove.bg
Bulk Image Resizer: https://www.birme.net/
How can I help in your endeavor?
In return, I would like to seek the community to share their images, creating a community-built facepack. This can be a newgen facepack for everyone, by everyone. The guidelines are simple:
The images should be 200x200 PNG files, with their background removed.
The image should not have any part of the hair or face cut-out of the image.
The images should be uploaded in the appropriate ethnicity folder following the NewGAN structure. This resource can be helpful if you are confused: https://customercare.23andme.com/hc/en-us/articles/212169298-23andMe-Reference-Populations-Regions
The images should be renamed to "[Ethnicity]xxxxxx", it is ideal to increment by the last image made.
The files will be saved in this Drive folder for everyone to contribute and download: https://drive.google.com/drive/folders/1VnKQlloIuny4JXP2VTqDbIjpZwibu5Cr?usp=drive_link
If you have any other questions, feel free to ask.
Happy New Year to all of you.
Regards,
gedmski
Over the past several months, I have dedicated myself to perfecting a revolutionary AI-driven solution – the FutGAN. In my relentless pursuit, I encountered numerous challenges while crafting a new-generation facepack for FM. While initial attempts yielded some successes, some persistent flaws and issues demanded a comprehensive solution. Today, I am thrilled to unveil the FutGAN – a cutting-edge Stable Diffusion model meticulously trained on FIFA player images. This breakthrough tool is a game-changer for managers seeking unparalleled realism and diversity in their gaming experience. Through trial and error and previous attempts that resulted in unsatisfactory or labor-intensive outcomes, FutGAN promises to redefine how we approach newgen face creation. What sets FutGAN apart is its remarkable ability to deliver realistic and unique faces, ensuring that newgens no longer appear repetitive or overly cartoony. Thanks to its versatile prompt-based generation, managers can now customize every aspect of their newgen players. Whether it's long or short hair, clean-shaven faces, full beards, or even neck tattoos – FutGAN empowers you to essentially give life to your newgen-infested squads. The link to the model can be seen here: https://civitai.com/models/245295 How do I create FutGAN faces? Generating these images may seem complicated at first, but here is a guide to help you start: There will be a Create button on the right side of the model. This will be the first step of your creation process. A new page on the left will now pop up, with details on how to generate your data. For this LoRA model, I recommend using epiCPhotoGasm. This checkpoint model provides realistic images with only short prompts, unlike other models that need longer prompts. However, you can use other models such as RealisticVision or epicRealism, you probably need to modify other details such as CFG and steps (will be provided for epiCPhotoGasm for this guide). The recommended weight for this LoRA is 0.85-0.95; anything lower will generate images from epiCPhotoGasm (too generic), and anything higher will provide deep-fried images. Luckily, we don't have to provide long prompts for epiCPhotoGasm, so a prompt including age, nationality/ethnicity, hairstyle, facial hair, and other characteristics can be easily added. To produce consistent results, add "plain white background" at the end of the prompt. Note: if you want to generate faces for players ages 19 or below, it is advised to use 20yo instead; new community guidelines have been implemented that prohibit the use of certain keywords. You don't need to do plenty of work for the negative prompt. Just copy this text instead: white hair, turban, cap, headwear, hat, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime), text, 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, UnrealisticDream Generally, it is not necessary to tinker with these settings, but I will give a brief description of these and recommended values. For this model, a square aspect ratio should be used. A CFG scale is a value that determines how much of the prompt will be followed; a lower value gives you a more creative and loose image and a higher value gives you more precise and generic (other times, deep-fried) images. I recommend using values between 4.0 and 7.0. DPM++ 2M Karras is the best sampler for this model. I don't see any evident changes when I set the step number, so any value is fine. The general rule is lower values give you less-effort images and higher values take too much computing power and time just for one image, I'd recommend values between 15-30. I would set a random seed for each generation just in case I'm not satisfied with the produced image. If you want to retain a consistent face and change facial characteristics for a specific player, you can input the UID of that player as a reference. VAE is sometimes baked into realistic models, so you probably don't have to add that. You can choose to have 1 image to 10 images, and there you have it! FutGAN faces. Other tools that can be useful to you: Background Remover: https://www.remove.bg Bulk Image Resizer: https://www.birme.net/ How can I help in your endeavor? In return, I would like to seek the community to share their images, creating a community-built facepack. This can be a newgen facepack for everyone, by everyone. The guidelines are simple: The images should be 200x200 PNG files, with their background removed. The image should not have any part of the hair or face cut-out of the image. The images should be uploaded in the appropriate ethnicity folder following the NewGAN structure. This resource can be helpful if you are confused: https://customercare.23andme.com/hc/en-us/articles/212169298-23andMe-Reference-Populations-Regions The images should be renamed to "[Ethnicity]xxxxxx", it is ideal to increment by the last image made. The files will be saved in this Drive folder for everyone to contribute and download: https://drive.google.com/drive/folders/1VnKQlloIuny4JXP2VTqDbIjpZwibu5Cr?usp=drive_link If you have any other questions, feel free to ask. Happy New Year to all of you. Regards, gedmski
