4K production test — identical recipe to hero pilot v1
Re-ran 3 of yesterday's heroes through the exact logged method — marketing_studio_image, identical prompt, identical reference packshot — changing only resolution: 1k → 4k. All three first-take, no drift. Output is true 4096×4096 (26–30 MB), well past Shopify's zoom spec. The reflective kettle (chrome + blue floral enamel) and the transparent strawberry glasses — the two hardest fidelity cases — came through clean. This locks the production model: it was always marketing_studio_image, and it scales to 4K at 4 cr/image.
Strawberry Old-Fashioned glasses
3L stovetop whistling kettle
Enameled cast iron round casserole
1K vs 4K — what changed
| Spec | Hero pilot v1 (yesterday) | 4K production (today) |
|---|---|---|
| Model / recipe | marketing_studio_image | marketing_studio_image — identical |
| Output size | 1024 × 1024 | 4096 × 4096 |
| Credits / image | 2 cr | 4 cr |
| Gen time (each) | ~30–60 s | ~3–4 min |
| First-take success | 31 / 31 | 3 / 3 |
Takeaway: the $1k full-catalog model stands exactly — it was priced on this model at 4 cr/4K. Quality at production resolution matches (arguably beats) the pilot. The only real cost remains QA review time, not credits. Local 4K files: work/higgsfield_test/hero_4k_test_20260613/. One watch-item: the casserole's embossed lid text reads "JS GOURMET" but AI-embossed text should always be verified per-SKU (or banded in-theme) before it ships.