Planet Plop
You can visit Planet Plop here. It’s just a sketch at this stage and will evolve over time.
Music and songs are created in collaboration with my friend Glenn Rogers.
From my perspective, I’m using it to explore and develop an emergent array of AI tools and techniques that can be used for a wide array of applications, whilst having a bit of creative fun. I’ll add some technical notes and thoughts here as I progress.
Planet Plop Microcosm
This was a fun experiment using SDXL with Deforum to generate synthetic DICOM data for MicroCT scans of imaginary microorganisms of Planet Plop. The volume data is visualised in Drishti and a narration provided by our science communicator, who bears an uncanny resemblance to a well-known naturalist and presenter from Planet Earth – but who obviously has far more refined sartorial sensibilities. V.O. was developed with ChatGPT 4, facial construction with SDXL (entirely through text prompts!) and animation with Sad Talker and Bark.
I initially set up Automatic1111 Stable Diffusion WebUI, managing the Python envs through Anaconda. But since that can take a lot of management, it’s much easier to play around with ‘AI package managers’ like Pinokio and Visions of Chaos. These are both excellent for a dealing with this rapidly evolving software environment and also for setting up local training systems if you have the compute available.
In terms of workflow SDWebUI is robust and intuitive, and has a rapidly growing ecosystem of extensions. For a more programmatic approach ComfyUI is the way to go – it’s far more configurable and can be chained into the kinds of workflows that I am interested in – so that’s where I’m focusing my attention for the time being. Hypernetworks and LoRAs (e.g. either self-trained or via online resources like civit.ai) enable the development of bespoke toolchains for rapid visual development and iteration – it’s fascinating stuff. Pairing these with OpenAI GPTv and similar APIs or local LLM installs (using e.g. Llama2 and vector databases like Pinecone) might ~in theory~ enable the development of iterative agents for complex visualization, programming and animation tasks. But that is for hand-wavy future research at this stage.
I’m especially interested in exploring how these systems can be interfaced via the lingua franca of python with interactive systems such as Unreal Engine, Nvidia Omniverse, Cesium and real-world observational datasets using gaussian splatting, NeRFs and satellite data deep learning. There are simply so many possibilities it can seem somewhat overwhelming, so constraints driven by the objective of specific projects is a good way to start – having said that, my imagination is running riot with all the things that could be done – some crazy mash-ups in this amazing new world that is just beginning to come into view. Of course, there are going to be hiccups and problems along the way, including a lot of fascinating and important questions of the sort inspired by Chalmer’s ‘Reality+’. What sort of media landscapes are our kids going to live in? It’s going to be a very different world from the one I experienced as a child (like Noggin the Nog, by the late great Oliver Postgate) and not everyone is going to be human. A range of interesting utopias are far preferable to the alternatives.
