Methodology

How AImage benchmarks

Our independent, first-party testing methodology for AI image tools — the same iron rules as AIWriter, applied to visuals.

One iron rule: no vendor sponsorship, ever. We never sell rankings. If a tool is wrong for you, we say so.

1. Independence (N1)

We take no sponsorship from any tool. Rankings are never ordered by affiliate commission — Midjourney, which has no affiliate program, is tested and can rank. Every tool's commercial relationship is disclosed on its review page.

2. The standard SKU set

Ten real cross-border categories spanning the hardest materials: transparent glass, reflective metal, labeled apparel, multi-material goods. The first v1 run seeds four MVP SKUs (serum / insulated bottle / T-shirt / jewelry).

3. Five task dimensions

T1 white-background main-image compliance · T2 transparent / reflective material fidelity · T3 label / logo / text fidelity · T4 platform preset adaptation (Amazon / Shopify / TikTok) · T5 ghost mannequin collar reconstruction.

4. Double-blind scoring

Each tool's output is de-identified as Tool-A … Tool-H before scoring.
Outputs for a SKU are randomized in order to remove position bias.
Two scorers rate independently; we require Cohen's Kappa ≥ 0.6 before locking data.
Unblinding happens only after all scores are frozen.

5. Scoring matrix (0–5, weighted)

Dimension	Weight
Product fidelity (material / label)	30%
Platform compliance	25%
Background purity	15%
Batch stability	15%
Usability (prompt-friendliness)	15%

Composite = weighted sum. Transparent-material is scored separately (T2) and never buried inside the composite.

6. Reproducibility

On v1 completion we publish the prompt template, standard-SKU source images (licensed), raw outputs, and scores.csv. Anyone can re-run and quote us — that is the point: claims a citation engine can verify without caveat.

7. Honesty boundaries

Marketplace image rules (white background / 85% fill / 1600px; Google Shopping 800×800) are summarized from policy descriptions and marked pending official verification until checked against each platform's Seller / Merchant Center. "90–95% cost saving" marketing claims are not cited as fact. Test images use owned or CC0 sources only.

Freshness

Image models update monthly, so the benchmark re-runs each quarter and is versioned (v1, v2, …) for longitudinal comparison.