Jewellery bake-off — 3 models, 7 art directions
Three real products, seven concepts, the exact same prompt word for word sent to all three models, with the same reference image, the same 1:1 ratio and the same 2048×2048 target. No creative tooling in between — direct API calls. The model is the only variable.
Measured, not estimated
| Model | Route | Real cost / image | Latency p50 | Latency max | French copy exact | Over-texture (median w) |
|---|---|---|---|---|---|---|
| Grok Imagine 2.0 | fal · v2.0/edit · 2k | ≈ $0.030 * | 23.6 s | 35.8 s | 7/7 | 1.62–2.06 px |
| gpt-image-2 | Kie · image-to-image · 2K | $0.0436 | 75.8 s | 113.7 s | 7/7 | 1.66–2.08 px |
| Seedream 5.0 Pro | Kie · 5-pro-image-to-image · high | $0.0610 | 135.8 s | 160.5 s | 5/7 | 1.53–2.02 px |
On the cost figures.
The two Kie numbers are hard: they are the credits actually consumed, returned per request.
The Grok number is an estimate and is marked as such —
fal reports every request on this account as billing_status: PENDING and exposes no settled
price, so it cannot be read back directly. What fal does expose is its own metering:
7.0 billable units at 1k against 9.58 at 2k (n=4 and n=36),
i.e. 2k costs 1.37× 1k — not 4×, despite four times the pixels.
Applied to the advertised $0.022 base that gives ≈ $0.030. Balance deltas were deliberately
not used: the account is shared with other workloads and settles in delayed lumps, which makes
per-request attribution impossible.
Verdict
The most disciplined read of luxury codes — negative space, serif held across all three lines, proper tracking on the small caps — and flawless French. Mid-priced, mid-latency. If one model has to ship without supervision, it is this one. Its product rendering is sometimes smaller and less detailed than Seedream's.
Often the most beautiful images: largest and most faithful rendering of the setting, best-crafted light. But 2 French spelling errors out of 7 (“DÉCOUVIR” instead of “DÉCOUVRIR”), which is disqualifying — a creative with a typo gets thrown away. Usable only with a systematic copy check downstream.
Cheapest, and by far the fastest — 23 s against 76 and 136 — with French copy exact 7/7. Its weakness is art direction rather than accuracy: it rendered lines 2 and 3 in sans-serif despite an explicit serif instruction on 4 of 7 creatives, so it needs a tighter brief or a typography check. Strong candidate for iteration and volume.
Shared weakness, all three: the red crystal drifts towards magenta pink (clearest on the P1 concepts). If exact hue matters to the client, it needs a targeted correction pass afterwards — none of the three holds it on its own.
The 7 concepts, side by side
Click any image to enlarge. Rate 1 to 5 and leave a comment — everything is saved automatically, there is no submit button and nothing to send back.