Krea 2 Raw/Turbo generate AI images in 2 s; Nano Banana...

Why does speed matter for AI image generation? Because developers and enterprises now weigh latency against licensing constraints. Mid‑2026 benchmarks show a spread from half‑second outputs to seven‑second waits.

FLUX.1 [schnell] from Prodia tops the chart at 0.5 seconds, released under Apache 2.0 with fully permissive commercial terms. Z‑Image Turbo, hosted by Replicate and fal.ai, clocks in at 1.8 seconds but remains proprietary—users must sign active API contracts to monetize results.

Here’s the thing: Krea’s own Krea 2 Turbo hits a clean 2.0 seconds, offered as a hybrid of open weights and proprietary licensing, accessible via trial or API. It preserves style‑reference compatibility and LoRA support while employing Trajectory Distribution Matching to shave off precious milliseconds.

Midjourney’s v8.1 Turbo mode pushes generation to 3–6 seconds, demanding a paid subscription tier and higher credit costs. Black Forest Labs’ FLUX.2 [klein] models sit just under five seconds—4‑billion parameters at 3.9 seconds, 9‑billion at 4.6 seconds—both under permissive licenses. Microsoft’s MAI Image 2 Efficient rounds out the field at 4–7 seconds, proprietary and billed per use on Azure.

These figures lay out the trade‑offs between raw speed, openness, and commercial viability.