Example of the parametric user interface in artbreeder.

Notice that they work on a new version of this user interface.

It could be hard to play with parameters like race or gender if you don't have data about the person (log in to a customer account) or decide which range of values you need. You could encounter an ethical dilemma so maybe it's better to let the styleGAN user to generate several versions of a persona depending of the country, the market etc.

It is hard or impossible to inject parameters like "Eye color" to StyleGAN as it is more about transferring styles, not specific defined features. The current repo doesn't include any specifications to suggest that and I doubt it will in the future.

However, to reach your goal, you can do the following (not ideal):

1. Train several image classification models (for example one for eye color, one for race, etc)

2. Generate thousands/millions of images using StyleGAN.

3. Apply classification models on the outputs of #2.

4. Now you can give images based on user options.

Thank you Issa Annamoradnejad for your feedback.

On the project Artbreeder, they added a parametric interface to change valence of several parameters which are still manage by the GAN model based on dataset but I don't know how they make the link between the model and these paremeters/variables:

• less or more happy
• less or more Angry
• less or more blue eyes
• less or more female or male
• less or more etc.

0 is the initial state/value of a parameter on a model.

Where you generate a model you restart with 0 for this model and can change the values (-/+)

We can see it on this video demo or play with the artbreeder site (funny activity) https://twitter.com/i/status/1185522974847393794

I added an example (you need to create a free account in artbreeder to edit picture) of a face I generated with white vs. black characteristics.

I'll install the package to work on it.

This paper is insteresting to add control on the latent spaces of the GAN.

GANSpace: Discovering Interpretable GAN Controls: Preprint GANSpace: Discovering Interpretable GAN Controls

Deepai: https://deepai.org/publication/ganspace-discovering-interpretable-gan-controls

PaperWithCode: https://paperswithcode.com/paper/ganspace-discovering-interpretable-gan

Videos:

https://www.youtube.com/watch?v=jdTICDa_eAI

https://www.youtube.com/watch?v=f4IHRcMXZM4

GANSpace is a technique for discovering interpretable controls for Generative Adversarial Networks (GANs) by performing Principal Component Analysis (PCA) in activation space.

Paper: https://arxiv.org/abs/2004.02546,

code: https://github.com/harskish/ganspace

Figure 1: Sequences of image edits performed using control discovered with our method, applied to three different GANs. The white insets specify the particular edits using notation explained in Section 3.4 ('Layer-wise Edits').

Another type of solution to generate pictures from words as parametric data this time.

New preprint, "Words as Art Materials: Generating Paintings with Sequential GANs", is now available online! Preprint Words as Art Materials: Generating Paintings with Sequential GANs

You can try the demo: https://painter-ai.com

pdf: https://lnkd.in/eyABDTs github: https://lnkd.in/eTqfAYk

In this study, Azmi investigated the generation of artistic images on a large variance dataset. The contributions can be summarized as follows:

• Created a new keywords-to-painting dataset. Dataset contains 3492 images with 1746 digital art images and 1746 classical painting images. There are 738 unique keywords. (This dataset is publicly available. You can get the download link from the github page.)
• Used keywords for image descriptions, where word order is insignificant and word relations are more obscure unlike full sentence descriptions, leading to more artistic image generations.
• Designed a sequential GAN architecture that handles both word vectors and real-fake image pairs.
• Employed a mixture of optimization techniques (Wasserstein loss, spectral normalization, and minibatch discrimination) to train all the networks in a stable manner and efficiently.

I found a good corpus of projects based on GAN with these services and APIs to generate faces:

Rosebud AI: AI generated models for marketing

https://app.generative.photos/

Generated Photos: beautiful datasets and api

https://generated.photos/datasets

https://generated.photos/api

Demo Ganvatar: there are Semantic Vectors to explore latent spaces

https://ganvatar.com/?ref=producthunt

Unique AI-Generated Faces Placeholders Sketch Plugin: tiny Sketch plugin that generates unique AI faces – data content placeholders – using thispersondoesnotexist.com

https://gumroad.com/l/thispersondoesnotexist

Development of "whole body model automatic generation AI" which generates whole body image of non-existent person. Use GAN to create realistic faces of personas is just a small step...

In this article: "I Am a Model and I Know That Artificial Intelligence Will Eventually Take My Job",

https://www.vogue.com/article/sinead-bovell-model-artificial-intelligence

I found a reference to Datagrid, a Japan company which uses generative adversarial networks (GANs), to create digital models which offer a vast array of posing options that mimic exactly what we do in e-commerce and commercial modeling. Datagrid: https://datagrid.co.jp/en/

The German e-commerce giant, Zalando SE, has published research papers on this technology :

Preprint Generating High-Resolution Fashion Model Images Wearing Custom Outfits

Conference Paper Generating High-Resolution Fashion Model Images Wearing Custom Outfits

Other paper: Semantically Multi-modal Image Synthesis

https://seanseattle.github.io/SMIS/

TL-GAN: transparent latent-space GAN, by Shaobo GUAN

A good three-week project: "Draw as you can tell: controlled image synthesis and edit using TL-GAN". Generating custom photo-realistic faces using AI. Controlled image synthesis and editing using a novel TL-GAN model.

Resource lists:

• Blog post expaining the motivation, architecture and results is posted on this Medium link: https://medium.com/p/d170b1b59255
• Slides presentation explaining the core ideas of this project are available at this Google Drive link: https://docs.google.com/presentation/d/1OpcYLBVpUF1L-wwPHu_CyKjXqXD0oRwBoGP2peSCrSA/edit#slide=id.p1
• A video presentation on YouTube: https://www.youtube.com/watch?v=O1by05eX424
• An interactive demo can be found in this Kaggle notebook: https://www.kaggle.com/summitkwan/tl-gan-demo

Core ideas:

• This project provides a novel method to control the generation process of a unsupervisedly-trained generative model like GAN (generative adversarial network).
• GANs can generate random photo-realistic images from random noise vectors in the latent space (see stunning examples of the Nvidia's PG-GAN), but we can no control over the features of the generated images.
• Knowing that the images are determined by the noise vector in the latent space, if we can understand the latent space, we can control our generation process.
• For a already well-trained GAN generator, I made its latent space transparent by discovering feature axes in it. When a vector moves along a feature axis in the latent space, the corresponding image morphs along that feature, which enables controlled synthesis and edit.
• This is achieved by leveraging a coupled feature extractor network (a CNN here in this demo, but can be any other CV techniques), which enables us to find correlation between noise vectors and image features.
• Advantages of this method over conditional GAN and AC-GAN:
• Efficiency: To add a new controller of the generator, you do not have to re-train the GAN model, thus it only takes
  0 votes 0 thanks

Build an app to generate photorealistic faces using TensorFlow and Streamlit (https://www.streamlit.io/)

We’ll show you how to quickly build a Streamlit app to synthesize celebrity faces using GANs, Tensorflow, and st.cache.

https://towardsdatascience.com/build-an-app-to-synthesize-photorealistic-faces-using-tensorflow-and-streamlit-dd2545828021

This tutorial is a good way to test an prototype quickly with GAN with controls on the latent spaces parameters.

Hello,

Teaching persona classes for students and professionals of museology, I received very mixed feedback on using pictures of real people in general. Many students/participants claimed that pictures would draw attention away from the persona features and convey prejudices and stereotypes. Instead, some of them preferred drawings in black & white without much details. However, this is only personal experience, I have no data to support these claims. I wonder how others think about this.

Hello Werner Schweibenz

produce stereotypes of real persons is one of the risks when we use personas. Using representation like realistic photography could let people think that this representation is "THE user or customer". And of coure we could have gender or racial representation issues which is a serious issue in some museums or part of the world.

For example, In Guadeloupe, West Indies, we have Mémorial ACTe Museum "to contribute to the construction of a collective memory of slavery and to finally healing its abhorrent wounds and hurts, through historical awareness, values of tolerance and contemporary creation." http://memorial-acte.fr/the-founding-acte-memory-holds-the-future

As the local population is mixed and that the Museum is about slavery, caribbean area and colonial history, use only a kind of persona (white or black etc.) will be a bad idea. So it's probably better to use:

• a panel of representations (like a matrix of people generated by exploring the latent space of racial types axis)
• or a very simplistic design which merge differents persons profiles
• or no representation at all

I encountered this kind of issues with personas designed:

• with marketing team with data and informations not updated over time (we know tht they change over time)
• for/with technical team which doesn't really want to work with a persona or UX-centered approach
• by UX designer to try to identify the main profiles of users implicated in a project or everyday (more like a program)

From last year, I work on a Data Hub Project (in a Data Roadmap where I included User / Customer Experience) and started by a typology of persons (Physical and/or Legal) to design the project not on stereotypes but on role (more Activity-centered ergonomics than UX-centered). After, with data and dynamic personas generation, it' possible to classify persons (thounsands if you have data) under each role or profile.

Roles can help to limit biases and trigger process design in organisation (in the top menu of the Memorial ACTe Museum, they added profiles entries: ADHÉRENT / FAMILLE / ENSEIGNANT / GROUPE / ETUDIANT / TOURISTE / PROFESSIONNEL ET ASSOCIATION)

I think that based on roles, we can create another granular level as cohort over the persona level and under the persona we add the natural (measured) or synthetic (calculated / aggregate) persons.

Level 1: Roles of Physical or Moral persons level

|____ Level 2: Cohorts of persona level (exploration of latent space on several axis)

|_______ Level 3: Natural or Artificial persona level (collection of data about agents splited by persona journey)

|__________ Level 4: Real person/agent/operator level (qualitiative and quantitative methods)

I don't limit experience (even we think about experience as observed behaviour of agent) to human and extend it to robots (animated objects or like agents in multi-system agent) and animal (being living) to be consistent with reality. As methods for evaluating long-term user experience like UX curves allow average values, you can use individual measurements on Level 4 aggregated on level 3 to calculate metrics for Level 2 and 1.

With this hierarchy, it's easier to design person profiling as natural/artificial (good to integrate AI), real, individual or cohort entities (fine for data and quantitative approach).

We have a candidate to explore and interpret the Latent Space of GANs for Semantic Face Editing.

InterFaceGAN - Interpreting the Latent Space of GANs for Semantic Face Editing

https://www.arxiv-vanity.com/papers/1907.10786/

Interpreting the Latent Space of GANs for Semantic Face Editing

https://paperswithcode.com/paper/interpreting-the-latent-space-of-gans-for

Arxiv PDF: https://arxiv.org/pdf/1907.10786v3.pdf

[...] "In this work, we propose a novel framework, called InterFaceGAN, for semantic face editing by interpreting the latent semantics learned by GANs. In this framework, we conduct a detailed study on how different semantics are encoded in the latent space of GANs for face synthesis. We find that the latent code of well-trained generative models actually learns a disentangled representation after linear transformations.

We explore the disentanglement between various semantics and manage to decouple some entangled semantics with subspace projection, leading to more precise control of facial attributes. Besides manipulating gender, age, expression, and the presence of eyeglasses, we can even vary the face pose as well as fix the artifacts accidentally generated by GAN models. The proposed method is further applied to achieve real image manipulation when combined with GAN inversion methods or some encoder-involved models. Extensive results suggest that learning to synthesize faces spontaneously brings a disentangled and controllable facial attribute representation"

Vidéo: https://www.youtube.com/watch?v=3dGKxdGk3dw

Preprint: Preprint Interpreting the Latent Space of GANs for Semantic Face Editing

Interesting approach to avoid ‘spatially entangled changes’ by editing attribute and latent code.

Style Intervention: How to Achieve Spatial Disentanglement with Style-based Generators?

Preprint Style Intervention: How to Achieve Spatial Disentanglement w...

Generative Adversarial Networks (GANs) with stylebased generators (e.g. StyleGAN) successfully enable semantic control over image synthesis, and recent studies have also revealed that interpretable image translations could be obtained by modifying the latent code. However, in terms of the low-level image content, traveling in the latent space would lead to ‘spatially entangled changes’ in corresponding images, which is undesirable in many real-world applications where local editing is required. To solve this problem, we analyze properties of the ‘style space’ and explore the possibility of controlling the local translation with pre-trained style-based generators. Concretely, we propose ‘Style Intervention’, a lightweight optimization-based algorithm which could adapt to arbitrary input images and render natural translation effects under flexible objectives. We verify the performance of the proposed framework in facial attribute editing on high-resolution images, where both photo-realism and consistency are required. Extensive qualitative results demonstrate the effectiveness of our method, and quantitative measurements also show that the proposed algorithm outperforms state-of-the-art benchmarks in various aspects.

pdf: https://arxiv.org/pdf/2011.09699.pdf

abs: https://arxiv.org/abs/2011.09699

Laurent Berry our study evaluating the applicability of artificial images for personas has just been published: Article Using artificially generated pictures in customer- facing sy...

Thank you Joni Salminen

I almost finish to read it.

The potential generalisation of your findings about "persona profiles which did not decrease the scores for authenticity, clarity, empathy, and willingness" are a first step and relevant for future projects where generate artificial faces for virtual (GUI, NUI) or physical (Object or Robot) natural user interfaces will be required.