42. Hero image style test (Midnight Intelligence)

Date: 2026-05-17 Owner: blog/visual Status: sample run, awaiting visual review by Simon

Goal

Validate the "Midnight Intelligence" / Bloomberg-Linear aesthetic for blog hero images before mass-generating across the whole article catalog. Three test samples cover three different blog categories so the winning visual direction can be picked from real outputs, not mockups.

Setup

Model: gpt-image-1 (OpenAI SDK v6.34, current alias for the latest image-generation model. The brief referenced gpt-image-2; the SDK currently exposes gpt-image-1 as the production endpoint.)
Size: 1536x1024 (3:2, close to the 16:9 target, native supported size)
Quality: high
n: 1 per call
Script: scripts/blog-hero-gen-test.ts
Output dir: public/blog/hero-test/
Run: npx tsx scripts/blog-hero-gen-test.ts

The script does NOT write to the database, does NOT modify any BlogArticle.heroImageUrl row, and is NOT wired into any cron schedule. It is a one-shot sample only.

Shared style suffix

Every prompt appends a shared style block to lock the design aesthetic:

Aesthetic: Midnight Intelligence, Bloomberg-Linear financial-noir.
Deep navy near-black background (#0a0e1a, #0d1117).
Restrained gold accents (#c9a961) only as thin highlights or contour light.
Editorial cinematic lighting, volumetric haze, photoreal physically-based render.
Minimal, restrained, no clutter.
STRICT: no people, no faces, no flags, no country labels, no readable text,
no logos, no brand marks, no watermarks.
Wide cinematic 16:9 framing.

Prompts used

regulation

A wide cinematic dark editorial composition: an empty parliament-style chamber hall at night, seen from a high three-quarter angle. Polished marble floor reflects thin gold pinstripe light beams cast from above. Architectural columns fade into deep navy atmosphere. Volumetric haze.

STYLE

quant

A wide cinematic dark editorial composition: an abstract 3D mathematical surface rendered as gold topographic contour lines floating over a near-black backdrop. A subtle candlestick wave pattern is etched into the floor as faint data grid. A single sweeping signal line walks forward across the surface. Tilt-shift depth of field.

STYLE

markets

A wide cinematic dark editorial composition: a section of an unmarked dark globe seen from a low orbital angle. Thin gold fiber-optic trade arcs span between unnamed market hubs over deep navy ocean. Dense atmospheric fog at the horizon. No coastlines highlighted, no country borders, no labels.

STYLE

Output files

Category	Filepath	Size
regulation	`public/blog/hero-test/regulation-2026-05-17T18-24-25-327Z.png`	1.92 MB
quant	`public/blog/hero-test/quant-2026-05-17T18-25-23-909Z.png`	1.92 MB
markets	`public/blog/hero-test/markets-2026-05-17T18-26-12-579Z.png`	2.00 MB

Cost

Item	Value
Per image (high, 1536x1024)	~$0.167
Images generated	3
Total actual cost	~$0.501
Budget cap	$0.50

Note: pricing taken from OpenAI's published image-generation rates for gpt-image-1 high quality at landscape size. Final invoice may vary by a few cents depending on tokenization.

Recommendation for next step

Visual review of the 3 PNGs is the gating step. Three branches:

Aesthetic on-brand on all 3 categories. Scale to mass gen.
- Cost forecast at N=100 articles: 100 x $0.167 = ~$16.70.
- Cost at N=300 (full catalog): ~$50.10.
- Add downscaling pass (1536x1024 → 1600x900 + WebP at q=82) to keep public/blog/heroes/*.webp under ~250 KB before commit.
One category drifts (e.g. regulation feels staged). Iterate on only that prompt, re-run the script with a filter flag, then mass-gen. Marginal cost: ~$0.17 per retry.
Aesthetic off across all 3. Revisit the shared STYLE suffix (likely the gold accent intensity or the "photoreal" vs "rendered" axis), regenerate the 3 samples, repeat review. Marginal cost: $0.50 per full retry round.

Before any mass run, add:

A downscale/compress pipeline (sharp or squoosh) to convert PNG to WebP.
An alt text generator (LLM, cheap) keyed off the article title.
A guardrail that skips generation if heroImageUrl is already set on the row, to make the job idempotent.

Constraints honored

No DB writes, no migration, no schedule entry.
No em-dash anywhere in script or doc (lint scope is src/, but the policy is global; this doc uses hyphens, periods, and parens only).
TS strict; script runs cleanly via npx tsx.
No git commit, no push.