Should I use a pro AI model for every image or only for final selected shots?

No — reserve the pro model for final selected shots only. Run all exploration, drafts, and concept testing on a cheap tier like Nano Banana 2 Lite at 3 cents per image, shortlist the winners, then re-render only those through Nano Banana Pro. At 1,000-image volume, Lite is four times cheaper than Pro with no loss to your final output.

Use a two-tier workflow: generate volume on the cheap tier, escalate only your selects to the pro tier. Nano Banana 2 Lite runs at 3 cents per image — 1,000 images for $30 — generates 2.5x faster than Nano Banana 2, and costs one-quarter of Nano Banana Pro per image (four times cheaper at 1,000-image volume). Google's own benchmarks show Lite outperforming Nano Banana Pro on text-to-image quality, so the old assumption that cheap means worse no longer holds for drafts.

The workflow in practice: generate 10–50 variations of each concept on Lite, review, shortlist 2–3, then re-render only the shortlist through Nano Banana Pro for final production quality. The pro pass is for shots that carry technical demands — fine detail, final-delivery resolution, hero placements — not for figuring out what the shot should be. Lite's 1K output resolution is a deliberate design trade-off that suits exploration; your finalists get the full-resolution treatment on Pro.

Why volume-first wins: at 3 cents and 4-second generation, you no longer budget individual generations — generating 10 options instead of one becomes the rational default. D2C advertisers apply this by testing 50 ad concept variants instead of 5 for less than the cost of a coffee, finding the winner, and cutting customer acquisition cost. Filmmakers apply it in pre-production: iterating on 20 mood boards used to be a budget conversation and is now practically free, so treat pre-vis frames and reference boards as effectively unlimited.

The one exception: go pro from the first generation only when a single image must succeed immediately — a hero or flagship shot where there is no exploration phase — or when the shot's technical demands (resolution, micro-detail) exceed what a 1K draft can even preview usefully.

Managing the volume: the real bottleneck at 20–50 generations per concept is not model speed but losing track of the creative direction you gave three images ago. The invideo agent holds persistent memory — load your character sheets, brand context, and creative direction once, and every subsequent generation in the session pulls from that context automatically while the invideo agent routes each task to the right model tier, Lite for exploration and Pro for selects. Since invideo carries the full image-model stack, including Recraft and GPT-Image-2 alongside both Nano Banana tiers, you never switch platforms to switch tiers.

Watch some of these to see what works for you:

only the ones that are chosen get escalated to Nano Banana Pro

— invideo's creative team