Aastha Valecha

“The model has never lost anything.” “It cannot navigate toward absence.” “Everything you see is inference.”

Concept

Dead reckoning is a navigation technique: given a last known position, a direction, and elapsed time, you estimate where you are now. It accumulates error. The further you travel from the last fixed point, the less certain your position becomes.

A language model navigating toward a human memory of loss is doing exactly this. It has a last known position — the statistical shape of how loss is described — and it moves from there. It has never lost anything. It cannot reach the original coordinate. What it produces is an accumulation of inference error that, paradoxically, often looks like grief.

The CVPR audience knows precisely why this happens. It still lands.

Installation

A participant types one sentence about something they have lost. The system generates a 4-frame archival strip navigating backward through latent space:

I PRESENT — Kodak Portra 400. Film grain, halation, chromatic aberration. Evidence.

II FIVE YEARS AGO — Kodachrome fade. Cyan drains first. Centre holds, edges blur.

III A DECADE BACK — Edward Hopper geometry. Hard light bar. Shadow diagonal. Stillness.

IV AT ONE LIMIT — Rothko colour fields. Two temperatures extracted from the scene palette.

Each strip is printed immediately at 300 DPI and mounted on the gallery wall. The participant’s sentence appears in the header. At the end of the exhibition, all strips are displayed together — a wall of other people’s inference errors.

Technical Implementation

This proof-of-concept renders all frame transforms procedurally using NumPy and PIL, with no machine learning dependencies, to demonstrate the visual pipeline. The full installation implementation would use:

Stable Diffusion XL (base) for photorealistic scene generation from participant text

DDIM inversion to extract a base latent from the present image

CLIP-guided traversal toward an “absence” embedding

~45 seconds per strip on a single A100 GPU

The procedural frame transforms — film grain, Kodachrome fade, Hopper lighting geometry, Rothko palette extraction — are fully deterministic and reproducible from the source code included as supplementary material.