Sources and Significance blogpost 2
AI as Meta-Medium: Medium Specificity vs Gesamtkunstwerk
Medium specificity landed in a world thatâs already multisensory and accelerated: purity survives less as an ideology than as a method. Our media environment is already multisensory and accelerated, so itâs unsurprising that hybrid forms command attention more easily than single-medium purity. Iâd argue that âpurityâ isnât dead so much as displaced: in AI pipelines like my ComfyUI character-sheet workflow, the site of specificity moves from materials (ink, film stock) to systems (prompts, seeds, LoRAs, samplers). Thatâs where style and authorship now crystallise. In this sense, AI behaves like a meta-medium: it can imitate many arts, yet each concrete configuration (graph + parameters) yields its own constraints the new ârules of the game".
Leaving the medium-specificity session, one line stuck with me: in a plural, entangled culture, âpurityâ survives less as an ideology than as a method. In AI pipelines like my ComfyUI character-sheet workflow, the site of specificity moves from materials (ink, film stock) to systems (prompts, seeds, ControlNets, LoRAs, samplers). Thatâs where style and authorship crystallise.
If Wagner fused arts on a stage, todayâs version often lives in latent space. AI collapses image, text and layout, so the question isnât âwhat medium is this?â but âwhich configuration did you choose and why?â
image: my ComfyUI graph for the character-sheet pipeline
I wanted to model a character from my favourite manhwa in 3D. Before sculpting, I needed a clean character sheet (front/side/back + expressions) to lock proportions and costume logic. Instead of drawing it from scratch, I designed a ComfyUI workflow that keeps style and anatomy consistent while giving me fast iterations I can take straight into Blender.
The media object: my comfyui character-sheet generator
The goal was to generate front/back/side and expressions for a character to be 3d modeled in blender.
This pipeline lets me âdrawâ by designing a system. Pose is blocked by OpenPose, mass by Depth, line discipline by LineArt. My âbrushâ is the graph; my âstrokeâ is a parameter change.
At a glance (mapped to my diagram):
Load models & resources - UNet / DualCLIP / VAE / LoRA.
Prepare character image input â reference portrait (identity & palette).
Define prompts - (a) character-sheet layout prompt, (b) style + quality prompt, (c) character descriptors.
Combine & condition - join strings â CLIP encodes; PuLidFlux / InsightFace help maintain face identity/embedding.
ControlNet operations - OpenPose (pose sheet), Depth (volumes), LineArt (edge discipline) via ControlNet SD3 Apply.
Generation parameters - latent noise + scheduler, sampler (DPM++ 2M Karras), and guiders.
Pose-sheet input - scaled to match canvas; guarantees orthographic consistency.
Image generation - SamplerCustomAdvanced â VAE Decode.
Upscale/Tiler and Grid assembly â print-ready sheet; save.
Composition from pose controls,
Form/lighting from depth control,
Line economy from line-art control,
Identity/style from LoRA + Flux/InsightFace,
Repeatability from fixed seed & sampler.
Medium specificity (today)
Classic modernism said each art should embrace its own limits (Greenberg). In my workflow, specificity lives somewhere else: in the configuration. The exact graph + weights + seed + sampler + ControlNet balance is what gives my sheet its look and reliability. Designing the pipeline is designing the form.
Wagner fused music, stage, architecture into a single experience. AI fuses text, image, layout (and soon motion/sound) inside one system unity conducted by a trained model. Itâs a kind of technical totality: immersion by generation. Thatâs powerful for production, but it can drift toward the same âAI sheenâ if I donât set deliberate constraints.
Good: rapid iteration, consistent multi-view outputs, accessible hybrid making.
Risk: style homogenisation and fuzzy authorship.
My guardrail: log seeds/samplers, limit variables per run, and keep a human pass when needed.
Notes from a hallway debate (friend ⢠me )
Friend: In a feed already multisensory and accelerated, single-medium purity struggles to hold attention. Hybrid orchestration (image + layout + motion/sound) matches how we now perceive.
Me: The 20th century saturated many âpureâ experiments. Mixing media creates fresh constraints (graphs, seeds, models) to explore. âMediumâ behaves more like a configurable system than a fixed material.
Lecturer: Pursuing purity is complicated under plural, entangled, stochastic conditions. Speak about specificity under hybridityâwhat a work does because of its setup, not its silo.
Working with ComfyUI moved my authorship from stroke to system. ControlNet blocks pose, the sampler is my âfilm stock,â the seed fixes repeatability. Thatâs where my voice lives what I lock and what I let vary. Purity isnât gone,I use it as a method (fix variables to isolate causation). The risk is homogeneity; the answer is intentional constraints and iteration notes so results are reproducible and mine.
Iâm now pivoting this workflow toward product animations. The goal is to keep the same pipeline logic (pose/volume/edge control, fixed seeds) but apply it to short, loopable promos that hold up under motion.
Bolter, J.D. (2023) âAI generative art as algorithmic remediationâ, IMAGE, 37(1), pp. 195â207.
Greenberg, C. (1940) âTowards a newer Laocoonâ, Partisan Review, 7(4), pp. 296â310.
(Reprinted in: Frascina, F. (ed.) (1985) Pollock and After: The Critical Debate. London: Harper & Row, pp. 34â46.)
Kay, A. (1984) âComputer softwareâ, Scientific American, 251(3), pp. 52â59.
Krauss, R. (2000) A Voyage on the North Sea: Art in the Age of the Post-Medium Condition. London: Thames & Hudson.
Manovich, L. (2001) The Language of New Media. Cambridge, MA: MIT Press.
Wagner, R. (1849) âThe Artwork of the Futureâ, in Wagner, R. (1895) Richard Wagnerâs Prose Works, Vol. 1. Translated by W.A. Ellis. London: Kegan Paul, Trench, TrĂźbner & Co., pp. 69â213.