Felix Brener
Contact
Felix Brener

Selected work04 of 04

Filmstock Responsiveness

A controlled study of how image models respond to filmstock, lighting and scan language in prompts, written up as a 31-page paper and replicated on a newer model.

Role
Sole author — design, analysis, paper
Year
2026
Status
Research
Stack
Python, PCA, MANOVA, PERMANOVA, ComfyUI
Code
filmstock-responsiveness-benchmark
A contact sheet of sixteen frames of a woman in a dark coat outside a corner store, labelled by model version, filmstock, lighting and processing.
Corner store, batch 1: Images 2.0 and 2.5 across Portra and CineStill, cool and warm light, clean and pushed scans.

Overview

Prompts for photographic images lean on the vocabulary of film: stocks, lighting setups, development and scanning. This study asks whether image models respond to that language, and by how much.

The first study compared ChatGPT Image 2 with Grok Imagine on a full-factorial grid — 288 normalised images, 195 extracted image features and 432 blocked prompt-response contrasts — analysed with PCA, MANOVA and PERMANOVA.

A September replication compared Images 2.0 with Images 2.5. Prompt-response distances were 24.0% lower for filmstock, 14.0% lower for lighting and 36.9% lower for development and scan wording (Holm-adjusted p < .001), and repeats of the same prompt were 41.5% closer together. The study measures these vocabularies only; it does not establish a general decline.

What I did

I designed the experiment, generated and normalised the image set, built the feature-extraction and statistics pipeline, and wrote the paper. The code is MIT-licensed; the data and writing are CC BY 4.0.

Numbers

normalised images in the first studyfilmstock-responsiveness-benchmark, README
288
image features extracted per imagefilmstock-responsiveness-benchmark, README
195
pages in the replication paperstudies/2026-09-images-2-5
31

Figures

Four bar charts of standardised feature distance for filmstock, lighting and scan wording, comparing Images 2.0 and Images 2.5 across primary, colour, contrast and texture, and structure features.
Prompt-response distances by feature family, Images 2.0 against 2.5.
A PCA scatter plot of 195 standardised image features, with confidence ellipses for ChatGPT Image 2 and Grok Imagine that overlap heavily.
PCA of 195 standardised features by model, first study.
Two bar charts: the ratio of factor distance to repeat distance by vocabulary, and mean same-prompt repeat distance for Images 2.0 and 2.5.
Repeat variation: the same prompt run again, measured the same way.

Related research

Criticality in sparse coding — a simulation study of phase transitions in a binary sparse-coding model, sweeping sparsity and overcompleteness across 600 Gibbs-sampling runs.

Heart-rate variability in concussion baselines — as a research assistant at Mount Sinai’s Abilities Research Center, I co-authored an abstract presented at the ACSM Annual Meeting in 2025 and published in Medicine & Science in Sports & Exercise.

A heatmap of susceptibility across sparsity and overcompleteness, bright where the model approaches its critical boundary, with the mean-field prediction drawn as a dashed line.
Susceptibility across sparsity λ and overcompleteness ρ, from the sparse-coding study.

Links