[07/14]
[ ]
GEN-AI RESEARCH X NOWHERE-NOWHEN [2022-PRESENT]
[ ]
PROJECT TYPE [DESIGN RESEARCH] MACHINE LEARNING [RUNWAY-ML GEN-1, DEPTH RENDER X OPTICAL FLOW] PROJECT OUTPUTS [DIGITAL ANIMATION-8K-PX]
[ ]
GEN-AI RESEARCH X NOWHERE-NOWHEN IS AN ONGOING PRACTICE-BASED RESEARCH INTO SYNTHETIC FORMS OF DESIGN INTELLIGENCE USING SELF-REFLEXIVE DESIGN METHODS. THE CORE EMPHASIS IS ON THE USE OF ARTIFICIAL INTELLIGENCE TO INVESTIGATE HUMANS AS THE MOTHER-SHIP OF ‘ANTHROPIC BIASES’ WHICH PERMEATES, THE NEURAL ARCHITECTURE OF THE EARTH’S BIOSPHERE. THE ANTHROPIC BIAS IS A SAMPLING BIAS, IN WHICH SAMPLES COMPATIBLE WITH THE SPATIOTEMPORAL POSITION OF HUMAN OBSERVERS ARE DEEMED REPRESENTATIVE OF THE UNIVERSE, ENTIRE. IN CONTRAST TO TURNING-OUT BIASES IN EMERGENT MODELS, AS OFT-CITED ALIGNMENT STUDIES PROMOTE, THE CORE AIM IS TO MISALIGN NEURAL REASONING FROM ANTHROPOMORPHIC TRAJECTORIES TOWARDS REASONING FROM NOWHERE AND NOWHEN, THE LATENT SPACE OF PURE NOISE AND SYNTHETIC FORMS OF PROBABILISTIC REASONING––MADE AMENABLE IN GENERATIVE AESTHETICS.
THE OUTCOMES INCLUDE NON-STANDARD WORKFLOWS FOR LARGE LANGUAGE MODELS, AND CONTENT-GUIDED VIDEO SYNTHESIS AND DIGITAL ANIMATION AND IMAGE SYNTHESIS WITH DIFFUSION MODELS. THE ANIMATIONS EXPLORES PROMPT-GUIDED DIFFUSION MODELS (RUNWAY-ML GEN-1) TO RE-COMPOSITE EARLIER ANIMATION WORKS: POST-URBAN TRAUMATICS (2015) & UNRELIABLE NARRATION (2016). THE DIFFUSION MODELS DE-LAMINATE THE DATA FROM THE SEQUENCE FOOTAGES INTO INDEPENDENT 'BANDS', RANGING FROM DEPTH CHANNELS TO GLOBAL SIGNED DISTANCE FIELDS. TOGETHER WITH AI-GENERATED AUDIO FEEDBACK, THE EXTRACTED BANDS INFORM SEGMENTATIONS, SHADER OPERATIONS, VIRTUAL CINEMATOGRAPHY & AESTHETIC FLOW IN TOUCH DESIGNER FOR GENERATING COUNTERFACTUAL FOOTAGE SYNTHESIS.
THE IMAGES EXPLORE GENERATIVE TRANSFORMER-BASED DIFFUSION MODELS (DALL-E II) WITH FOCUS ON IN- & OUTPAINTING, IMAGE-GUIDED CONTEXT EXTRAPOLATION & SEMANTIC LOSS IN SPATIAL EXPANSION & FEATURE PREDICTION. THE PROMPT ENGINEERING INCLUDES TYPE-TOKEN OPTIMISATION THROUGH LARGE-LANGUAGE MODES (GPT-3, 4) WITH EMPHASIS ON SEMANTIC SCORE. THE DESIGN SEARCH STARTS WITH A REAL-WORLD IMAGE AS A CONTEXTUAL SEED, TO FROM WHERE A TEXT-GUIDED PROMPT ADDS IN-STEP NOISE, GUIDING THE MODEL TOWARDS AN OPEN-ENDED OUTCOME. IT IS AN INVERSE PROCESS, AS OPPOSED TO DE-NOISING IMAGES OUT OF LATENT SPACE, THE MAIN FOCUS IS ON ACHIEVING DE-COHERENCE & SEMANTIC LOSS. THE RESULTING IMAGES ARE UPSCALED USING BICUBIC INTERPOLATION.
[ ]
SELECTED SAMPLE: SCROLL DOWN.
