How do you use cheap AI models for drafts and premium models for final output?

Run a two-tier escalation workflow: generate every mood board, variant, and draft on Nano Banana 2 Lite at 3 cents per image, shortlist by concept fit, refine prompts while iteration is still cheap, then re-render only the chosen assets on Nano Banana Pro — the premium tier that costs roughly four times more per image at volume.

The workflow is one funnel with five steps: draft in volume on the cheap tier, shortlist, tune prompts while mistakes are cheap, escalate winners to the premium tier, and hold creative context across the whole session. invideo is an agentic video creation tool with all the current image and video models available, so the invideo agent can route each generation to the right tier without you switching tools.

1. Draft everything on the cheap tier. Use Nano Banana 2 Lite for mood boards, ad-concept variants, product mockups, and pre-visualization frames: 3 cents per image, 2.5x faster than Nano Banana 2, and 1,000 images for $30. Its 1K resolution is a deliberate design trade-off, not a quality failure — drafts are judged on idea, not texture — and at that price generating 10 options per concept instead of one becomes the rational default. As one invideo review put it: if generation is that cheap and that fast, you don't budget your generations anymore.

2. Shortlist by concept fit. Review the batch and pull the 2–3 frames that carry the idea — composition, concept, and read, nothing else. This is where volume pays: a D2C team can test 50 ad variants for less than the cost of a coffee, find the winner, and cut customer acquisition cost, instead of testing 5 on a constrained budget.

3. Refine prompts on the shortlist — still on the cheap tier. Every wording, framing, and style correction costs 3 cents here instead of a premium render. Lock the prompt completely before anything escalates, so the premium pass is a re-render, not another round of experiments.

4. Escalate only the chosen assets to Nano Banana Pro. Route final production frames and technically demanding shots to Pro; everything else stays on Lite. At 1,000-image volume Lite runs four times cheaper than Pro, so a 20-draft-plus-3-final pass costs a fraction of rendering all 20 on the premium model — while you evaluated 20 concepts instead of 3. Internal draft-review costs are effectively eliminated at Lite pricing.

5. Hold context across the volume with persistent memory. The bottleneck in high-volume generation isn't model speed — it's losing track of what you told the model three images earlier. Load character sheets, brand context, and creative direction into the invideo agent once; its persistent memory applies that context to every subsequent generation without rebriefing, and it routes each task to the appropriate model tier automatically.

The same split applies on the video side inside invideo: draft and review at a model's default 720p output, take the free 1080p upscale for approvals, and pay for the 4K render — priced at the cost of a full generation — only on locked finals. For filmmakers, the practical effect of tiering is that pre-production exploration stops being a line item: iterating on 20 mood boards used to be a budget conversation and is now practically free.

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