Nano Banana 2 Lite + Gemini Omni Flash: A Cost-Efficient Storyboard-to-Video Stack

AI video budgets are often wasted before animation begins. A team asks a video model to explore composition, wardrobe, location, camera angle, and action simultaneously, then pays for several moving clips that were conceptually wrong from the first frame. A storyboard-first stack moves those decisions into cheaper still images.

Google’s June 30, 2026 releases make that pattern especially clear. Nano Banana 2 Lite—served through the API as gemini-3.1-flash-lite-image—targets fast, inexpensive 1K image creation. Gemini Omni Flash turns selected ideas into short 720p video with sound. Using each only for the stage it handles well can make creative iteration both cheaper and more legible.

Why still images should carry early uncertainty

A storyboard is not decoration. It externalizes the shot list so a director can reject a weak composition before motion, audio, and render time become attached to it. Reviewers can compare alternate framing, check product prominence, identify missing coverage, and decide whether a sequence tells the story.

The economical pattern is a funnel. Generate many inexpensive panels, narrow them through review, refine a small number of anchor frames, then animate only approved candidates. The funnel needs explicit gates; otherwise low image prices simply encourage an unreviewed flood of assets.

Define the brief before generating: audience, format, aspect ratio, visual style, subjects, continuity rules, required product details, prohibited content, and shot purpose. Every panel should answer one production question.

Nano Banana 2 Lite’s actual contract

Nano Banana 2 Lite is Google’s nickname for the generally available Gemini 3.1 Flash Lite Image model. The model card describes 1K image output, 14 aspect ratios, local editing, and a latency target below two seconds for typical generation. Actual latency depends on input and service conditions; the figure is a product target, not a per-request guarantee.

The 1K cap is appropriate for boards, thumbnails, layout tests, and many web assets. It is not a 2K or 4K master generator. Avoid enlarging a board and treating invented detail as approved product truth.

Local edits are useful for changing one element while preserving the rest of a panel: adjust a jacket color, move a prop, simplify a background, or clear space for copy. Save the original and edited image as separate versions with the edit instruction.

All images generated through Gemini include SynthID. Store the model ID and generation metadata as well, because an embedded marker alone does not explain the prompt, references, or editorial changes.

The consistency limitation matters

Google positions Lite as the fastest and most economical image model, but its image-generation guide says it is not optimized for multiple reference inputs or sequential multi-turn editing. That is a meaningful boundary for storyboards, where the same character, product, and location may recur across many panels.

Do not solve this by hiding dozens of prior images in an ever-growing conversation and assuming the model remembers exact details. Create a visual bible: approved character descriptors, wardrobe swatches, product geometry, palette, lens language, environment rules, and one canonical frame. Repeat the critical constraints in a compact template.

Generate panels in small groups and compare them side by side. When exact reference composition or multi-turn consistency is central, use a stronger Nano Banana tier for the anchor frame, then return to Lite for lower-risk variants. The cheap model should explore; it should not silently lower the continuity standard.

Design the shot manifest before animation

A production manifest turns the board into machine-readable state. Give every shot a stable identifier and record its narrative purpose, panel file hash, aspect ratio, intended duration, camera move, subject motion, audio cue, required continuity, and approval status.

Keep generation prompts separate from delivery instructions. “Slow dolly in while the subject turns” belongs to the animation request. “Place this after shot 04 and use the approved voice track” belongs to orchestration. Mixing them makes revisions harder to audit.

The manifest also prevents duplicate spending. If a render request times out, the application can check whether a job already exists for that shot and panel hash instead of submitting again. A conversation transcript is not a reliable production ledger.

Move only approved frames into Omni Flash

Gemini Omni Flash supports image-to-video within its public-preview API. It produces 3–10 second video at 720p and 24 fps and attempts to generate sound. An approved storyboard panel gives the model a strong initial composition; the prompt can concentrate on motion, camera, environment changes, and audio.

Use one visual beat per clip. State what should remain fixed, what should move, the direction and pace of movement, the camera behavior, and the intended end state. Avoid conflicting adjectives and impossible timing. For dialogue, keep the line short enough for the selected duration.

Review the first and last frames as well as the middle. A clip can begin attractively and end with a changed product, face, or background. Inspect temporal stability, hands, text, logo shapes, motion direction, and whether generated sound is usable.

Omni can also use reference and editing modes, but the current preview has limits: no audio references, voice editing, extension, interpolation, or multiple-video reasoning. It should not be treated as an automatic long-form assembler.

A transparent cost example

Google’s current pricing lists Nano Banana 2 Lite 1K output at an effective $0.0336 per image in standard processing and $0.0168 in batch. Its underlying rates are $0.25 per million input tokens and $30 per million image-output tokens. Input costs and retries are additional.

Twelve standard storyboard panels therefore cost about $0.40 in image output. Google lists Gemini Omni video output at $17.50 per million tokens and 5,792 tokens per generated second, or approximately $0.10 per second at 720p. One eight-second shot is about $0.81 in video output. Together, that illustrative funnel is roughly $1.21 before inputs, alternative video takes, storage, and finishing.

Batch image pricing is useful for non-interactive exploration, while standard calls suit a live review loop. Do not generate twelve clips merely because twelve panels exist. If reviewers approve three, animate three.

The more meaningful metric is cost per accepted shot. Track how many panels, video attempts, edit rounds, and reviewer minutes were needed. A more expensive anchor image may reduce enough failed animation to lower total cost.

Quality gates for the complete funnel

At storyboard review, check composition, continuity, rights, product accuracy, and whether the shot adds information. At animation review, check prompt adherence, temporal coherence, duration, frame rate, audio, and safe content. At assembly review, check sequence, pacing, transitions, streams, branding, and delivery codec.

Separate creative rejection from technical failure. A valid but boring clip should not be retried automatically. A transient API error may be retried with the same idempotency record. A changed prompt or panel is a new creative version.

Keep original sources and generated assets immutable. Derivatives should reference their parents, not overwrite them. This makes it possible to revisit the sequence when a model preview changes or a stakeholder requests a new cut.

Deterministic finishing through separate Medux jobs

The generative stack ends with selected video files. Medux is not a built-in component of Gemini, and no native integration is claimed. It can serve as a distinct finishing service through authorized MCP workflows after the creative outputs are locked.

If Omni’s generated sound is rejected, use the Claude MCP audio-track tutorial to attach an approved voice or music file. For exact branding, the Codex MCP logo guide describes a separate overlay operation. Assemble approved shots in manifest order with the Claude MCP video-merge workflow.

Give each operation its own input hash, parameters, idempotency key, task ID, status, output, and validation result. Do not ask an agent to remember whether it already submitted a job. Poll by the returned ID, bound retries, and require approval before publication.

Validate that the selected logo and audio are the approved versions, not temporary storyboard assets, before either write operation begins.

This division keeps the funnel economical and reversible. Nano Banana 2 Lite explores visual ideas, Omni Flash spends video compute on selected frames, and Medux applies defined finishing operations. The savings come not only from lower model prices, but from deciding earlier, recording state, and refusing to animate uncertainty that a storyboard could resolve.