Antonio Olivera La Rosa Antonio, The easiest way to begin is a Google "Images" search of, “database of faces”. It opens a wide variety of faces and sources. Usually from top of head to the neck. They vary in dimension, resolution and usage rights. Many are faces that were collected from public websites and social media. If needed, you can use photo software to make uniform dimensions and similar tone/color

In addition; there’s open source photo databases of faces that were collected from public websites and social media. Plus stock photo agencies and the Getty Museum will sell you a collection of diverse faces.

RE: “Artbreeder”…this is “morphing software. With your input specs it combines two images into one. I believe you want a AI software** that can create faces from text input. For example; “create 40 faces of male and female American teenagers”.

** Or commercial design / photo software that has built-in text based AI tools.

==============

I found the following resources but can’t verify their usefulness. I also didn’t take the time to look at the face databases, see if they're accessible online and/or by downloaded. I’m leaving it to you to explore them :

1)__ Princeton UNIV (in USA) has a collection at

https://libguides.princeton.edu/facedatabases

2)__ You can look at iMerit face database at

https://imerit.net/blog/5-million-faces-top-17-free-image-datasets-for-facial-recognition-all-pbm/

3)__ From the Center for Lifespan Psychology, Max Planck Institute for Human Development, Berlin, Germany you can look at “Faces” at:

https://faces.mpdl.mpg.de/imeji/

4)__ A face database by MIT Univ (in USA) at

https://web.mit.edu/emeyers/www/face_databases.html

5A)__ If you’d like to design the search tool for images on internet - a tool, Milvus is useful. Its a Vector Database tool. With a free get-started mode at

https://milvus.io/

5B)__ A similar model is Sefik Ilkin Serengil. Its described as “Large Scale Face Recognition with Pinecone Vector Database” and it at;

https://sefiks.com/2021/05/19/large-scale-face-recognition-with-pinecone-vector-database/

6)__ From the journal, PLOS, an article describing the “The Bogazici face database: Standardized photographs of Turkish faces” with supporting materials”. It’s at this link;

https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0192018

7)__ Don’t know if this is helpful but used for game design….”a collection of character face presets created, to be used with “expired6978 's LooksMenu mod”. Link is;

https://www.nexusmods.com/fallout4/mods/13961

==================

8)__ From GitHub….This also requires exploring the databases to determine if the content is available. It shows the “AT&T Database of Faces” (from mid 1990's) and includes info for other collections; MNIST, CIFAR-10, CIFAR-100 and CALTECH101. But its unclear if access is for downloading images. Link is at:

https://git-disl.github.io/GTDLBench/datasets/att_face_dataset/

9)__ The Chicago Face DataBase. This has a number of related collections. You might have to request access. The link is;

https://www.chicagofaces.org/

Lee Clawson I really appreciate your response. What I'm looking for is AI software that can create composite faces from the input of existing faces. Specifically, creating realistic faces from a large number of faces. The objective is that the resulting face is representative of the population used.

Antonio Olivera La Rosa --- Antonio, from your recent description I have more questions. For example; do you want to input a small number of representative faces and have software that "automatically" generates many variations __or__ input the same example faces and add descriptions of (text input) the parameters (visual details of their faces) that are representative of the population ?? I believe text input can be accomplished with variations of CHATgpt-4 and a few of the Adobe applications. The morphing of images (Artbreeder et al) is the basis of the idea. Typically it generates one image. You'd have to program a "morph" to then morph itself multiple times. Is this closer to what you want ? A related question; can existing databases of faces from specific cultures function just as well....or...is generating the faces a criteria ?? One remote possibility is to look at this as a "photo-retouching" problem. Instead of making a single solution you're asking to see 100's.

Lee Clawson thanks for your time. I want to input a large pool of example faces and merge all the input faces into one "typical" face, representative of that population. The final result should look like this:

https://theonlinephotographer.typepad.com/the_online_photographer/2013/09/averaged-faces-of-various-nationalities.html

However, instead of using standarized photos (that is, similar in format as in passports), my objective is to use naturalistic photos, that is, photos of real people who are not expressly posing for the photo, and therefore, may differ in certain aspects (emotional expression, for example).

Hope it is clearer now.

Best,

Antonio

Antonio Olivera La Rosa Antonio, I'm working on a video and won't have time to read the link until tomorrow (I'm familiar with The Online Photographer"). ------ But, the link to FaceResearch.org and (my alternate) FaceLab.org both fail. Until I get time to read this I suggest doing a search for "Face Research"... and... "Face Research on ResearchGate". The latter returns numerous links. Possibly you can find others working in this area. Also, take a look at FigShare.org (located in London). It's a strange looking web site but they mention (towards the bottom) using "Psychomorph" or "WebMorph.org". I assume these are the "morphing" tools they use (???). __ Morphing software raises with a hardware question; while I don't know whats needed for your project do you have access to computers capable of image processing and morphing ???

Thanks a lot for your time Lee Clawson . I am familiar with the more "classic" morphing software used in face research. However, creating composite faces with these software typically requires the use of standardized (non-naturalistic) faces, which is not the case for me. That aspect makes them not a very good option for what I am looking for. All roads may lead to AI solutions

Antonio, I read the Online Photographer article; I assume its set of faces that you wanted me to see, because I didn't get much out of the text or comments. Two items are worth looking at, (1) Francis Galton technique to discover (among other tasks) the "criminal face" and (2) using the "Median Filter" (see this link

https://en.wikipedia.org/wiki/Median_filter).

Here's where I am with this; (1) assume you need a computer capable of image processing and (maybe) AI. (2) My experience with "morphing" software is limited to pre-AI. But I ever encountered any need to use standardized imagery. The software was design to output a single image at a default dimension, resolution and 8-bit color. It had, compared to today and AI, limited parameters to chose from. The AI image generating tools I've used also output a single image but via text offer a much wider number of parameters. Creating photos and video using on-line tools are less effective than desktop tools. Where I am the tools are being offered via subscription. And quite costly for bigger projects. Adobe and similar companies are gradually offering AI tools built-in to their existing applications. You input text on your own images. And their tools are "trained" on their own libraries. Your university might have a subscription to the Adobe collection of applications. If so, I would start these tools. And use them to learn how to think about the project. With more sophisticated tools you should be able to ask the tools to look at specific features of an image and specify the what shoud occur. However, from the links and a mere cursory understanding I believe you're going to want tools that are trained to see the elements that make up the face. Or an AI tool that lets you train it on features that are essetial to your project. In this way I suggest collaborating with people that have the tools to analyze faces versus undertaking the development yourself. --- A large group of face morphing people work in game design and "special effects studios". They modify (train) standard AI tools on their visual needs. By the nature of their work they are familiar with writing code to "mode" the tools. ----- For example the traditional example of morphing is the face of a tiger and a male face. The basic tools let you "dial in the percent to mix them. The next level brings focus to eyes, noses, lips and chins. More or less the tools used by "plastic-surgeons" to et you preview your new nose. Today these have become "socially-accepted" AI tools. The next set are a variety of "pre-viz" tools, that is, pre-visualizing the scenes of film and animation. ----- If I was starting this project I'd look locally for others working in similar fields. Maybe in your university or others nearby. ---- I'll stop here. Feel free to add comments or try again to explain what I'm missing.........If I get some time I'm going to see what the London link is about__________

Lee Clawson thanks for your valuable feedback. Adobe AI tools looks like an interesting option. I´m also considering AI tools like StyleGAN, which may be a very good option:

https://github.com/NVlabs/stylegan?tab=readme-ov-file

What do you think?

Best,

Antonio

Antonio Olivera La Rosa Antonio, I’m working and I’ll explain the following later. -- I looked rather quickly StyleGEN; I like it. And it appears to be free. Like that too. - I use the Adobe tools. They’re very good but I’m not sure they will easily (I need to explain this too) generate a series of faces. I looked at the GitHub site for StyleGEN and have two questions; (1) You can download the code and compile it yourself. Are you or your colleagues OK with this ?? - (2) There’s little mention of computer hardware that will run this efficiently. I saw a vague mention of nVidea. They make GPU’s. Often tools like this are designed to use the GPU and often a specific model. Do you know what’s needed to run StyleGen ??? — I’ll write and say more about Adobe. And maybe find some installation notes for StyleGEN.

Thanks Lee Clawson for your time and extremely valuable feedback. For now, the priority is to have a solution that can serve as a starting point and StyleGEN seems to be a stong candidate. We consider the possibility of incorporating someone specialized in GEN IA if necessary. Best, Antonio

Antonio, You're welcome. --- I read, “incorporating someone specialized in GEN IA if necessary”. I think thats a great idea. Even if it’s for limited time. This stuff is new and hard to learn how to think about it. -- I’m not minimizing Adobe and similar products but for now the attached PDF gives more attention to StyleGEN. ——— The PDF has notes I collected about StyleGen and my experiences with Adobe tools. The notes regarding StyleGEN are described in more detail on the bottom of the “StyleGen” GitHub web site page at this link….

https://github.com/NVlabs/stylegan

Thanks a lot Lee Clawson ! Very useful tips. I think the project is already starting to take shape, I really appreciate your time and dedication.

Antonio Olivera La Rosa Antonio, As the project progresses; let me know what you use and the computer set-up

Antonio Olivera La Rosa .... Antonio, You may find some use for this... The link (below) is for the Kaggle dataset for faces. This is a part of a much larger dataset named “ImageNet”. -- Each require you to log-in and agree to their ”terms of use”

https://www.kaggle.com/datasets?search=faces

Thank you Lee Clawson , best.

Antonio Olivera La Rosa Antonio, I I haven’t looked into the following items**…but…my guess ?? they might be useful alternatives. The AI apps are; Cogvideo, Mochi, Kling and Hulio. **What prompted me to write is Kling and Hulio use “Low Rank adaption” (LoRAs ).

Low-rank adaptation (LoRA) is a technique for quickly adapting machine learning models to new contexts. LoRA helps make huge and complicated machine learning models much more suited for specific uses. It works by adding lightweight pieces to the original model, as opposed to changing the entire model. LoRA helps developers quickly expand the use cases for the machine learning models they build. –– More information on LoRa at this link;

https://www.cloudflare.com/learning/ai/what-is-lora/#:~:text=Low%2Drank%20adaptation%20(LoRA),without%20retraining%20the%20entire%20model.

Antonio Olivera La Rosa antonio, this article may be of interest as well. The link is;

Article Exploiting Feature Space Manipulation for Image Generation i...

Thanks a lot Lee Clawson ! LoRA looks like a promising alternative, we´re still in development phase so this information is quite valuable.

Best,

Antonio