Polyguard · Butyl Stainless Steel Tape · Image-model bake-off

Job-Site Photo — Model Test

Same real product photo + same scene prompt through four image models, to find which one reproduces the actual product (thick, dimpled stainless facer over tan butyl).

The target — the real product

Actual Butyl Stainless Steel Tape roll

The reference every model was locked to: bright silver metallic roll, visible stacked layers at the cut edge (thickness), dimpled grey tape strip.

Results — top picks first

gpt-image-2
E · gpt-image-2RecommendedMost photoreal + premium. Holds the silver metallic, stacked-layer edge, and dimpled applied strip. Newest OpenAI model.
gpt-image-1.5 high fidelity
F · gpt-image-1.5 · input-fidelity HIGHBest detailCrispest stacked-layer edge fidelity to the real roll. The high-fidelity setting is the key that was missing.
FLUX Kontext single
A · FLUX Kontext (edit)Very faithful (edits the real photo) but less premium lighting than the gpt-image‑2/1.5 renders.
FLUX Kontext multi
B · FLUX Kontext-multiFaithful, similar to A.
gpt-image-1
D · gpt-image-1Good baseline with the correct reference; the newer gpt-image models beat it.
Gemini nano-banana
C · Gemini (nano-banana)Scene onlyBest inventive scene, but flattens the product — least faithful on thickness.
Recommendation: gpt-image-2 (with gpt-image-1.5 + input-fidelity HIGH as the detail backup). Two fixes cracked it: (1) lock to the real roll photo that shows true form + thickness — not the smooth-foil or flat closeups; (2) use the newest gpt-image models with high input fidelity so the product is preserved, not reinvented. We'll produce the full photo set this way.

Bezar Video Studio for Polyguard · internal model test.