[Writing] AND ((((Sampling)))) (2) : LIAR
2024
This project is the culmination of a long exploratory phase, during which I mainly relied on the Automatic1111 graphical interface, a web interface widely used for Stable Diffusion, which offers more advanced features than basic consumer platforms.
The protocol implemented for this project entrusts autonomous algorithms with the choice of text prompts and the duration of the denoising process, which are the two functional pillars of the Stable Diffusion system.
With this work, I sought to test what would remain of my engagement with this system when I abandoned the illusion of a direct and linear causality between prompts and the images produced. The goal was to identify the extent to which a deliberate withdrawal of agency from these models allows new tasks to emerge on which the singular influence of the user can be shifted.
Ce projet synthétise l’aboutissement d’une longue phase exploratoire, au cours de laquelle je me suis principalement appuyé sur l’interface graphique Automatic1111, une interface web largement utilisée pour le modèle texte-image Stable Diffusion, qui offre des fonctionnalités plus avancées que les plateformes grand public de base.
Le protocole mis en place pour ce projet confie à des algorithmes autonomes le choix des invites textuelles et de la durée du processus de débruitage, qui sont les deux piliers fonctionnels du système Stable Diffusion.
Avec ce travail, j'ai cherché à tester ce qui resterait de mon engagement avec ce système lorsque j'abandonnerais l'illusion d'une causalité directe et linéaire entre les invites et les images produites. L'objectif était de dégager dans quelle mesure un retrait agentiel délibéré avec ces modèles permet de faire ressortir de nouvelles tâches sur lesquelles l’influence singulière de l’utilisateur·rice peuvent se déplacer.
Recursive process:
1. I started with a classic Gestalt optical illusion that plays on the ambiguity between the detection of the word “Liar”, in cursive, and that of a face – thus cultivating a sense of perceptive tension, from the outset.
2. I tried to minimize as much as I could my involvement in a desired input content, by allowing an algorithm to dynamically adjusts the required number of iterative denoising steps. This choice made the generation process relatively slow, while not being able to know in advance when an image will be outputted.
3. In parallel, I delegated to another algorithm the task of generating prompts at each stage. My influence then shifted to the intuitive weighting of certain sections of the prompt through a syntax the model can process: the parentheses increasing the weighting and the brackets decreasing it.
4. Each output image is then superimposed on the previous ones and the resulting layering becomes the input from which the algorithm generates a new prompt. As we can see, the initial optical illusion ended up in an almost shapeless stratification.
5. Likewise with the prompts: each textual input is added to the previous one to form an increasingly heavy and incomprehensible one.
6. The whole work was finally animated by frame interpolations to show how, the more the work amplifies, the more the recognizable figure blends into a surface made of tokens and visual patterns that are, not only unreadable to me, but also independent of strict auctorial authority on my part.
