vanitas_veritas

No water, no dream. A sink you can't wash your hands in: its water is a GPU dreaming vanitas still lifes without end, poured from the tap to the drain. By Katherine Titterton, San Francisco, 2026.

Three documents, each linking the others:

• the doc: the piece, how the water is made, the price; print it for a submission
• the build guide: parts, wiring, setup and upkeep
• the black paper: the thesis, where it goes, and texts and a rider for hosts and juries

The water itself, live: veritas.

Two pieces, one water

• vanitas is the sink: a freestanding stainless-steel sink with a self-portrait above it, designed and not yet built (first for tiat's Epistemologies of Slop, 2026). Real water lies over a screen in its bowl, and the tap runs; an old MacBook Pro under it dreams the picture.
• veritas is the same basin on the web, verse/sync.html. A hand in either moves the water in both, over public Nostr relays.
• The faucet handle is the tap. Hold it and generation stops, and every basin drains.

Kept small, kept running

Each choice was measured on an Apple M5 with 16 GB.

• 2.3 GB of GPU memory: the model in half precision; only the decoder in full, which keeps colours clean.
• Memory handed back after every dream and stretch: the process went from 10 GB to 5.8 GB.
• The upscaler in its own process, on the Neural Engine: Core ML holds Python's lock, so it cannot share one with the dreaming.
• No black frames: attention slicing turns fp16 dreams to NaN on Apple GPUs, so it is off, and an empty dream raises instead of pouring.
• It keeps flowing: a GPU error restarts from the last dream; a stalled upscaler gives a plainly enlarged frame and restarts; six frames wait in a queue poured every 2.5 to 3.5 s.
• Little kept: the last 60 frames on disk; prompts walked by a stride, never stored; the place in them survives a restart.
• Little shown: the spring serves only the page, the newest-frame note and the frames, never its key or its state.

Run it

On any Apple-silicon Mac:

caffeinate -i python3 code/sink.py            # veritas at http://127.0.0.1:8770/
caffeinate -i python3 code/sink.py --public   # and to everyone, through a Cloudflare quick tunnel

On the sink's own MacBook Pro, for good (the build guide walks through it):

bash code/setup_sink.sh     # packages, models, cloudflared, two test dreams
bash code/sink.sh install   # the spring and the full-screen page, at every login

The moodboard lives in source/moodboard/ and the Real-ESRGAN weights in code/models/ (both local only). --public writes each new tunnel address to verse/spring.json, so the plain veritas address finds the water. Midjourney pictures saved by hand to source/mj_inbox/ become the next dream.

Files

path	what it is
code/sink.py	the spring: dreams, walks, upscales, pours, serves veritas
code/upscale.py	Real-ESRGAN x4: Neural Engine for the spring, GPU for the loop
code/latent.py, code/harmonize.py	slerp, Catmull-Rom and the VAE's way in and out; the palette lock
code/setup_sink.sh, code/sink.sh	the sink's computer: install once, then run for good
code/morph_v44.py, code/morph_v45_sr.py, code/order_preview.py	the 2026 loop the spring grew from (rendered with the render-loop skill in .claude/skills/)
verse/sync.html	veritas: one page, WebGL2, no build step
,

Colophon

Stable Diffusion 1.5 with LCM-LoRA, Real-ESRGAN, stable fluids (Stam, 1999); Midjourney for the style, by hand. Built with Claude and Claude Code. Code under the MIT License; the images are the artist's.