GPP image-gen — v2 (model-tailored + reference-image test)

Two experiments: (1) Reference-image prompting — uploaded 3 of Justin's real photos to Gemini as anchors, asked for the same plumber in 3 residential scenes. (2) Model-tailored prompts — same 5 residential scenes as the v1 run, each prompt rewritten per model (Gemini: descriptive 8-segment structure, Grok: front-loaded + negative hack, OpenAI: concise positive-only). See what the prompting research actually buys.

→ Reference-image test → v2 3-way compare → v1 vs v2 delta v1 gallery (original) full photo library

Reference-image test — Gemini with 3 Justin anchors

Upload 3 real Justin photos as references + ask Gemini for the same plumber in a new residential context. Goal: preserve actual body/uniform/gear from his real photos while transplanting the scene.

Justin (from refs) under a residential kitchen sink

anchor: justin-064.jpg
anchor: justin-031.jpg
anchor: justin-044.jpg
GEMINI OUTPUT

Justin (from refs) at residential water heater install

anchor: justin-064.jpg
anchor: justin-031.jpg
anchor: justin-044.jpg
GEMINI OUTPUT

Justin (from refs) walking up suburban driveway

anchor: justin-064.jpg
anchor: justin-031.jpg
anchor: justin-044.jpg
GEMINI OUTPUT

v2 — 3-way compare (model-tailored prompts)

Same 5 residential scenes. Each prompt rewritten per model per the technique research.

Hero — plumber on suburban colonial driveway

slug: res-01-driveway-hero
Gemini 2.5 Flash Image v2
Grok Imagine v2
OpenAI gpt-image-1 v2

Detail — hands + wrench on angle valve

slug: res-02-kitchen-sink-wrench
Gemini 2.5 Flash Image v2
Grok Imagine v2
OpenAI gpt-image-1 v2

Trust — hand writing on clipboard at island

slug: res-03-clipboard-island
Gemini 2.5 Flash Image v2
Grok Imagine v2
OpenAI gpt-image-1 v2

Exterior — plain white van in driveway

slug: res-04-van-driveway
Gemini 2.5 Flash Image v2
Grok Imagine v2
OpenAI gpt-image-1 v2

Install — 50-gal water heater in closet

slug: res-05-water-heater-closet
Gemini 2.5 Flash Image v2
Grok Imagine v2
OpenAI gpt-image-1 v2

v1 vs v2 delta — side-by-side per scene

Same scene, v1 (original prompt) left column per model, v2 (tailored) right. Dimmer v1 tiles means "pre-research." See which scenes benefited from the rewrite.

Hero — plumber on suburban colonial driveway

Gemini 2 v1
Gemini 2 v2
Grok Ima v1
Grok Ima v2
OpenAI g v1
OpenAI g v2

Detail — hands + wrench on angle valve

Gemini 2 v1
Gemini 2 v2
Grok Ima v1
Grok Ima v2
OpenAI g v1
OpenAI g v2

Trust — hand writing on clipboard at island

Gemini 2 v1
Gemini 2 v2
Grok Ima v1
Grok Ima v2
OpenAI g v1
OpenAI g v2

Exterior — plain white van in driveway

Gemini 2 v1
Gemini 2 v2
Grok Ima v1
Grok Ima v2
OpenAI g v1
OpenAI g v2

Install — 50-gal water heater in closet

Gemini 2 v1
Gemini 2 v2
Grok Ima v1
Grok Ima v2
OpenAI g v1
OpenAI g v2

Nagatha · 2026-04-21 v2 · model-tailored prompts per research at workbench/notes/2026-04-21_image-prompting-research-gpp.md