Presentation Objects of Interest

Introduction

Aims of this workshop

1. To give you an insight into how we are working with AI image creation
2. Have a dialogue on possible intersections with your research.

Agenda

1. Background (Chr.)
2. AI image generation: techniques and infrastructures - evading the conventional platforms and capital interests (Nicolas & Pablo)
3. Conversation

The background (Knowledge Servers)

Within a larger project on digitization, democracy and citizenship (SHAPE), we run a project that not only questions the role of Big Tech and platforms in the construction of knowledge, but also how one might ‘delink’ from capital interests.

We try to speak from the point of view of what you might call grassroots’ internet culture. A culture that had control of the tools, infrastructures and other means for its own existence – what the anthropologist Chris Kelty once called a ‘recursive publics’.

History took us elsewhere. To 'client-server' (platform) relations (rather than peer-to-peer).

We speak of technical infrastructures as a “epistemic infrastructures” (borrowing the term from Celia Lury): An infrastructure for what and how we know things – and not just for how we find information, or use it in other ways.

Project image made with Stable Diffusion

Our questions

How does grassroots digital culture respond to this development? OR What does technological autonomy look like today?

… What are the strategies for culturally, democratically and environmentally sustainable knowledge infrastructures?

… What are the tensions within technologies?

… How do you provide alterantives/other services and infrastructures for communities?

Our methods

We collaborate with grassroots groups (such as collectives that run such services)

We share a history of being part of this culture ourselves

Our 'amateurism' perhaps even attempts to be ‘undisciplined’ – suspending disciplinary regimes (challenging the conditions under which knowledge occurs)

We run and build services, "knowledge servers":

… a server (within the university ecosystem)

… a library

… a wiki

… an etherpad

… a git repository

Portrait of a self-hosted server

How we proceeded with generative AI

Our outset has been Stable Diffusion - a different visual culture than, say DALL-E.

A journey/mapping of, guided by Nicolas.

Mapping 'objects of interest/necessity'

... Objects, not as manifestations of an ideal (an imaginary of autonomy), but the other way around: in our investigation we have found these objects and we try to draw a map of what autonomy looks like, rather than what it is.

... That is the objects that create the imaginary

... Not what autonomy is, but what it looks like under given conditions.

Resulting in a workshop/exhibition/catalogue, September 2025 (final outcome)

Sketchy map of our journey

There are things that show

... ‘pixel space’: Images we create or images that are used to create and annotate data sets.

There are things that don't show

… A 'latent space' of models that generate outcomes

… Software and people/communities that use and build on the software, to – for instance – support their own visual culture.

… A level of organization of these people (for crowdsourcing p2p, using virtual tokens)

… A material level of dependencies on units (such as the GPU, minerals, 'dead labour')

AI Horde image generation with a communal infrastructure

Why?

• Monopoly and concentration of power in platform capitalism
• The user is the product
• Governance
• Epistemic closure

Platform economy

Numbers of newly funded AI startups per country

Taller Estampa, map of generative AI

Software

Stable Diffusion page, screenshot

Stable Diffusion, training process

Stable Diffusion pipeline, Severine Dusollier

Network of actants, Stable Diffusion

---

Decentralisation

CivitAI - a platform

CivitAI is a highly popular platform/repository for distributing AI models and images. In our taxonomy, it's what we refer to as a space between the private and the commons. It's private in the sense that it's focused on individuals sharing their work (e.g., the creation of new models), and commons in the sense that most models can be reused and remixed. It's a repository whose primary purpose is to share resources.

CivitAI has its own currency ("buzz"). It can be purchased or earned by being part of the community (posting, recommending, commenting, etc.). It can be used to leave tips or to order certain models.

It is largely a cultural hub, with its own economy, rules, and aesthetic preferences.

While many, and the most popular models, recreate canon's of beauty and heteronormative standards, civitAI can be read as a hyper-specialisation space beyond big tech. This is where the "amateur modeler" can share their work.

AI Horde - a network

AI Horde is a distributed computing cluster system. Using open source and crowdfunding, it allows users to borrow and lend other computers with graphics cards to use generative AI models.

The system works by creating a directory of "workers" who are part of this network and distributing calls among themselves. In this sense, it is partially decentralised.

As a worker, you can generate "kudos," the exchange currency with which you can also generate images of other workers (more kudos = better and faster models).

Artbot - an interface

There are many interfaces for this "decentralised" network of "workers" (since the AI horde software/network is accessible through APIs, access can take many forms)

One is a web-based interface: artbot

Not only is there a complex system behind this interface, like with most generative AI interfaces (computational models based on large, obscure databases), but there is also a somewhat invisible intersection between the technical, the cultural, and the artistic (computational, computer science, the amateur modeler, the visual artist, and the average user).

Artbot not a plug-and-play interface either. While it is a suitable version in which programming knowledge is not required (as in older versions), it is not a blank, tame interface (think of Apple or chatGPT)

If you want to use the GPU in the room, click (as per the image below), and search for the worker called "xpablov117"

If you choose to use "our" worker, you will be limited to the following models (no LORAs):

• stable_diffusion
• Deliberate
• ICBINP - I Can't Believe It's Not Photography
• AbsoluteReality

(here is a list of all the models available in the network at this point: https://tinybots.net/artbot/info/models

About the NSFW issue within AI horde:

• https://www.patreon.com/posts/nsfw-question-72771484
• https://www.patreon.com/posts/blacklist-72890784
• https://dbzer0.com/blog/ai-powered-anti-csam-filter-for-stable-diffusion/

Configuring AI horde

(if we have time, we can talk/show the file configuration for AI horde)

Installing AI horde

Running AI horde

Dialogue

❧ What does it mean to make images in these ways?

❧ How does governance function in these projects?

❧ What is the relationship between the sharing ethos of these projects and their value generation?

❧ What forms do the economy, labor, and bodies take?

Resources

Startups funded 2013-2023:

• https://www.visualcapitalist.com/mapped-the-number-of-ai-startups-by-country/

Critical perspective:

• Generative AI ecosystem by Taller Estampa

AI and materiality

• Big data, and only getting bigger https://arstechnica.com/information-technology/2025/05/where-hyperscale-hardware-goes-to-retire-ars-visits-a-very-big-itad-site

Software

• DiffusionBee official website (MAC)
• Diffusion Bee GitHub, AGPL Free Software
• Draw things official website https://drawthings.ai/
• ComfyUI (all platforms), Git Hub, Source
• https://github.com/AUTOMATIC1111/stable-diffusion-webui, Diffusion WebUI aka Automatic 1111

Training LoRAs

• Manual Hugging Face
• Hugging Face LoRA dataset creation
• Adding a LoRA to generation pipeline
• Training LoRA in ComfyUI
• Essential to Advanced Guide to training a LoRA
• Making a LoRA is like baking a cake
• Fine tuning Stable Diffusion with LoRA