Prompt obedience, not polish: Luma Uni-1 Edit Max 69.7, Ernie Image Lora 56.5

In a head-to-head run published by runtimewire.com, Luma Uni-1 Edit Max scored 69.7 against Ernie Image Lora's 56.5 across eight image tasks ^[1], and took the task ledger 6 wins to 0 with 2 ties ^[2]. The aggregate spread of 13.2 points ^[3] is the less interesting number; a shutout on individual tasks is harder to explain away as house style.

The reviewer reports the result as a win at 99% confidence ^[4]. The method: eight tasks generated on the fly for the matchup so neither model could prepare, scored by gpt-5.4, with every task judged twice in swapped presentation order and every reported figure the average of both passes ^[5].

The named wins are about following instructions, not looking good. According to the review, Luma took perspective and scale with a clean one-point-perspective shot down an empty library aisle, negation by actually excluding the forbidden objects, exact counting with a true overhead seven-cup flat lay, and the Marshgate shelf tag by delivering the requested Swiss-style placard and line breaks rather than improvising ^[6]. It also handled the rainy citrus scooter ad better, landing closer to the specified compact electric scooter, branding and flat-vector treatment ^[7].

Ernie's problem, in the reviewer's words, is that it often looked good while being wrong ^[8]. Its library frame was attractive but put a large foreground book and table in the aisle, which breaks an empty-aisle brief ^[9]. The cozy reading nook ignored negation constraints and left visible wall art and shelving in frame ^[10]. The shelf tag drifted into a supermarket scene and collapsed the typography ^[11]. The recurring shape of the near-misses was the same: right count but wrong camera angle, good mood but weaker context, polished composition missing the specified object form ^[12].

The two ties read as genuine draws rather than narrow losses. On hands and anatomy the judges split on which image better conveyed the bracelet-tying action, which the reviewer takes as neither model fully locking the assignment ^[13]. On missed-train commuter, Ernie was stronger on narrative props and context while Luma was stronger on subtle facial emotion and cinematic portraiture ^[14]. The kettle task also split across passes, one preferring Ernie's simpler stovetop product read and the other Luma's cleaner palette and studio restraint ^[15]. Since five Luma wins and both ties are named individually out of eight tasks, the kettle is the sixth win on the two-pass average, and it is the one the reviewer flags as not effortless ^[16]. The final call is that Luma is the model to trust when the prompt actually matters ^[17].

What to watch: eight tasks and one judge model is a thin base, so the question is whether Ernie's counting, negation and typography failures reproduce on a larger set or were sampling luck. The failure modes here are the cheap ones to verify in-house, since object counts, exclusion lists and text blocks are checkable without a judge model at all. Also worth watching is whether a single order swap is enough to neutralise position bias when two of eight tasks flipped between passes ^[15][13].