About                   Projects                   Research                Consultancy














Polymorph II is a result of research into multisensory, distrubuted AI working across different forms of matter, using the indeterminacy of complex physical systems to fine-tune generative AI models and produce emergent outcomes in immersive media.



















It was produced using a recursively fine-tuned Stable Diffusion model and a real-time audio synthesis model (RAVE), alongside distorted steel plates functioning as sensors and sound resonators. In its data collection phases, thin strands of conductive ‘hair’, moving with shifts in air currents in the room, operated as sensors, together with live camera and audio input.










The data which drives these audiovisual outcomes is the result of the confluence of small changes in air current, the movement of bodies, and fluctuating electromagnetic interference entangled with fine-tuned generative AI models. Transforming across formats and forms of matter, the dataset generating the visible and auditory components of the work is merged with the environment of the data collection architecture.
FULL  VIDEO, VisLAB, London, 2025


This technique has enabled ‘leaps’ and ‘leaks’ - discontinuous transfers across the AI environment, where data formats and structural layers intersect at multiple temporal scales and across material components, generating differential intensities across sensing, auditory, and optical processes that do not resolve into a single hierarchy.

Sensing, auditory, and optical elements operate simultaneously as inputs and outputs, producing a continuously reconfiguring manifold of feedback relations.













    In this iteration, Polymorph II integrates a RAVE audio model, a fine-tuned Stable Diffusion model, two steel plates (2)  that function simultaneously as sensors and sound resonators, and a suspended conductive steel thread (1) that shifts closer to and further from a curved metal sheet. These changing proximities register as continuous variations in signal, producing data that is fed into TouchDesigner.

    Subtle changes in air currents, body movement, and electromagnetic interference modulate the system. The steel plates translate vibrational activity into both acoustic output through surface resonators (3) and data input, while the conductive thread introduces low-intensity fluctuations through its movement relative to the metal surface. These streams condition the activation of the generative models.

    The Stable Diffusion model generates images that are projected onto the metal sheet and surrounding walls. The sheet’s curvature (2) (8) redistributes the projection, producing a shifting optical field that is captured and reintroduced into the system through cameras (10). The resulting forms, transient, creature-like figures, act as markers of the system’s state, continuously shaped by incoming data.

    Generated outputs are stored in local folders and used to further fine-tune the model, allowing the dataset to expand through its own activity. A small monitor displays a live imprint of these changes (6), indicating shifts in operational phase.

    A microphone records the composite acoustic environment (5). This signal conditions the RAVE model, which generates sound played through speakers (4), feeding further variation back into the system.

    Sensing, generation, and modulation remain entangled, with signals continuously circulating across material, acoustic, and visual states.















The physical structures and embedded sensors tuned the system to its own material conditions, so that its behaviour shifted in response to micro-variations in air currents and its own internal dynamics.

These changes registered as subtle but persistent modulations across the monitors.















Curved metal sheets within the installation distorted reflections of both the generated outputs and the camera feeds. These distortions were captured and fed back into the system, contributing to ongoing adjustments of the model.




The curved metal sheets acted as reflective transformation surfaces within Polymorph’s wider feedback ecology. They bent, fragmented, and displaced projected images and live camera feeds, allowing signals to return to the system as altered input. Their curvature changed spatial proximity and continuity: elements that were separate in one pass could become adjacent in reflection, while minor surface distortions could accumulate through repeated recapture.

This made the sheets central to the project’s investigation of unstable distributions. Signal, image, vibration, movement, and light were not held in separate channels, but continually redistributed across material surfaces, sensors, and generative processes. Inspired by cephalopod skin patterning and avian murmuration, the sheets tested how large-scale organisation can arise from local deformation, imitation, and feedback, without requiring a central controlling image or fixed sequence.














As data passes through different formats and material states, the dataset generating the visual and auditory components of the work gradually expands, merging with the environment in which the system operates.



Systemic phase change - Polymorph’s activity maps


In Polymorph, a phase change refers to a reorganisation of the system’s dynamics that emerges within the interaction of many coupled processes rather than from a single controlling variable.

The system operates as a multivariate, adaptive setup in which sensors, models, and material feedback continuously influence one another. Under these conditions, small variations can accumulate and push the system across critical thresholds, leading to a change in how its activity is coordinated.
At first, the maps functioned as traces of ongoing activity. As the system evolved, they began to reveal recurring regimes: periods where the system’s behaviour stabilised around particular patterns, followed by systemic reorganisation and transitions into different configurations. 

These regimes formed through the system’s own dynamics as it moved between more and less coherent states.
A phase change, in this context, describes the transition between such regimes: a shift in the organisation of interactions across the system, where new patterns of coherence emerge and previous ones dissolve. 

The maps make these shifts readable. They indicate when the system reorganises itself, and when distributed interactions settle into a different configuration of stability.

Datasets and multiagent complex dynamics





This dynamic diagram demonstrates an alternative approach to datasets within a complex AI environment by drawing on graph-theoretical modelling. It begins with the conventional distinction between vertices and edges, where vertices correspond to datasets and edges to the connections between them, and then departs from static graph structures by treating connectivity as variable and continuously reconfigurable. As data is redistributed across the system, connections intensify, bifurcate, or disappear, producing ongoing internal reorganisation.

The image series presents this process across three spatial dimensions and one temporal dimension. Filament-like edges register increasing complexity as datasets split, merge, recluster, and regroup. As these shifts accumulate, the graph is reconfigured from within: its topology, understood here as the overall organisation of connectivity across the system rather than any local configuration, changes, and with it its density, distribution, and internal spacing. Nodes that initially appear relatively uniform, corresponding to datasets of comparable size, gradually develop into irregular bulb-like forms as redistribution proceeds unevenly across the system. In the final stage, the process becomes recursive, as the environment no longer serves only as the source from which data is drawn but is itself reshaped by the changing organisation and behaviour of the datasets.

The work was made by first modelling datasets and their connections as a graph, and then allowing changes in the data to drive changes in the graph’s form. As datasets split, merged, or changed their relations, these variations were translated into differences in spacing, density, node deformation, and connective filaments. The sequence of images records successive stages of that transformation. The diagram is therefore designed, but not freely invented: its forms were generated through an indexical process constrained by the behaviour of the datasets and by the parameters chosen to register that behaviour.


 Polymorph proposed a distinctive approach to datasets in a complex system. In many conventional AI workflows, the model is specified in advance as an architecture, and the dataset is assembled during the data collection phase to provide the statistical material through which the system is trained. During training, the model’s parameters are iteratively adjusted to minimise a defined loss function, allowing it to become responsive to recurrent patterns, correlations, and structural tendencies present in the data. These regularities are distributed across the parameter space, so that the system produces outputs in accordance with the statistical constraints it has internalised. Geometric descriptions can be applied to these learned structures, particularly in relation to embeddings or high-dimensional organisation, though the training process itself consists in continuous optimisation rather than the extraction of explicit forms.

From there, the model may undergo fine-tuning. In simple terms, fine-tuning involves additional training on a smaller, more specific dataset after the main training phase has taken place. This process is typically carried out under more controlled conditions, often with lower learning rates or partial constraint of the network’s parameters. It enables the model to become more specialised, more stylistically consistent, or more responsive to a particular domain, without requiring the entire system to be retrained from the beginning. A broadly trained image model, for example, can be fine-tuned on a narrow visual corpus so that it develops a stronger sensitivity to a particular class of forms, textures, or motifs.

It is broadly correct to say that, in standard machine learning pipelines, datasets have their most direct and explicit role during training. That is the stage at which the model is shaped through exposure to examples. At the same time, the dataset continues to operate beyond this phase, as its statistical structure persists in compressed form within the trained parameters. Every subsequent generation remains conditioned by this prior exposure. When the model responds to a prompt, it samples from distributions that were formed during training, and those distributions carry forward the constraints established by the dataset. In diffusion-based systems, this becomes particularly clear, as generation proceeds through an iterative denoising process that guides stochastic initial conditions toward regions of higher probability defined by the training corpus.

Within this context, Polymorph introduces a shift in how datasets are treated. The dataset becomes part of an ongoing process, rather than remaining confined to an initial preparatory stage. Outputs can be reintroduced as inputs, environmental signals (collected by Polymorph’s sensors)  can enter into the system’s operation, and subsequent retraining can alter the distribution from which future outputs are drawn. This kind of recursive and materially situated data circulation extends beyond standard one-pass training pipelines and aligns  with forms of continual or adaptive learning, while introducing additional layers of environmental coupling. The dataset, in this configuration, should be understood a component within a system that continues to reorganise over time.



Datasets are first modelled as vertices connected by edges, establishing a sparse graph structure in which relations appear  stable, discrete, and evenly distributed.





As data is redistributed across the system, previously uniform nodes begin to expand, deform, and differentiate, registering changes in dataset size, composition, and proximity..
With further splitting, merging, and reclustering, connections thicken into filament-like bundles, indicating a more complex internal organisation and a growing density of interdependence across the graph.

In the final stage, the graph registers recursive system-level reorganisation, as data circulation feeds back into the environment and modifies the conditions shaping subsequent transformations. The characteristic flaking marks the detachment of residual elements from earlier dataset configurations, as components no longer stabilised within the dominant structure separate and begin to consolidate into emergent subsidiary clusters.








Sequential outputs from Polymorph arranged as a matrix of generated forms. Each image functions as a temporary state within a continuous morphing process, where individual outputs do not remain isolated artefacts but pass into one another through gradual shifts in texture, topology, colour, and density. Across the matrix, local variations begin to produce larger recurring tendencies


Polymorph outputs first dispersed through standard t-SNE similarity-based grouping, then re-clustered through flocking behaviour












Plot of image data across two arbitrary semantic dimensions.

Polymorph’s continuous training outputs are not unambiguously separable, producing images that produce strong activations across disparate dimensions, here for classes “nipple” and “pretzel”.


Plot of t-SNE mapping of Polymorph outputs. To illustrate the various phases of Polymorph, the outputs from different stages of its activity were classified using a 152-layer ResNet model and mapped with the t-SNE algorithm to reduce dimensionality. The grouping of outputs expresses various functional shifts while maintaining cross-generational similarity.

The ResNet classification was performed using a pre-trained model from PyTorch. Each generated image was transformed to produce an output embedding of size 1,000, with each element corresponding to a distinct semantic category from ImageNet (for example “magpie”, “coffee mug”, or “screwdriver”).

The t-SNE algorithm reduces the >1,000-dimensional data to a two-dimensional plane while reflecting similarities between elements. The formation of clusters results from shared features in the input data.




Systemic taxonomies: ”Creatures”



















To observe the system as it gradually increased in complexity, a set of entities was generated that functioned as discrete markers of its behaviour. These evolving forms registered shifts in sensor input and internal dynamics, taking shape through the interaction between environmental data and the system’s ongoing processes.







Their morphology was continuously modulated by incoming signals and by the outputs of the fine-tuned models. The creatures, along with other generated forms, were produced within the system and then fed back into it as part of the ongoing training process.





































VisLAB, London, 2025







Experiments in which forms generated by the Polymorph system were placed in an environment with underwater visual patterns, reflecting the conditions of the initial training data.




Entities generated by the system were translated into three-dimensional forms, whose behaviour within the synthetic environment was continuously captured, converted into data, and fed back into the system to inform subsequent transformations.