A production multimodal content pipeline for GPT Image 2, Nano Banana 2, and Veo 3.1 Lite

Decision summary

Build around the content mission rather than a permanent model choice. A campaign may need on-brand stills, text-heavy product graphics, short motion variants, or an editable sequence. Route each job by required fidelity, reference behavior, resolution, latency tolerance, safety profile, and review burden.

The launch posture should be approval-gated: brief, generate candidates, persist originals, run technical and policy checks, obtain a responsible human decision, publish through an authorized account, and record the result. Fully autonomous external publishing combines probabilistic generation with brand, rights, platform, and relationship risk before the team has an operating baseline.

SOURCES: [2] OpenAI · [4] Google AI for Developers · [5] Google AI for Developers

Current image model map

OpenAI's current default image model is GPT Image 2, model ID `gpt-image-2`, with snapshot `gpt-image-2-2026-04-21`. It supports image generation and editing through the Images and Responses APIs. The Images API is suited to one-shot generation and editing; the Responses API supports conversational, editable experiences. GPT Image 1.5 is deprecated.

Google's “Nano Banana” name now covers four Gemini native image models. Nano Banana 2, `gemini-3.1-flash-image`, is the go-to balance. Nano Banana 2 Lite, `gemini-3.1-flash-lite-image`, is optimized for speed and cost but limited to 1K and is not optimized for complex multi-reference or sequential editing. Nano Banana Pro is `gemini-3-pro-image`; the legacy model is `gemini-2.5-flash-image`.

SOURCES: [1] OpenAI · [3] OpenAI · [2] OpenAI · [4] Google AI for Developers

Current Veo 3.1 Lite limits

The exact preview model is `veo-3.1-lite-generate-preview`, introduced in the Gemini API changelog on March 31, 2026. It accepts text or image input and produces one video per request with native audio. Supported durations are 4, 6, or 8 seconds; output is 720p or 1080p, with 1080p requiring an 8-second duration.

Lite does not offer 4K, the separate up-to-three-image `referenceImages` ingredient workflow, or video extension. It does support an initial image and first-to-last-frame interpolation. Google documents request latency from about 11 seconds to six minutes, says English is the evaluated prompt language, and notes that a seed only slightly improves repeatability rather than guaranteeing determinism. Generated videos are retained for two days, so a production pipeline must copy successful output to owned storage immediately.

SOURCES: [5] Google AI for Developers · [6] Google AI for Developers

Reference pipeline and operating model

Store the creative brief as structured data: mission, audience, channel, claim boundaries, required elements, prohibited elements, reference assets, aspect ratio, accessibility needs, and approval owner. The router selects an approved model and snapshot or preview based on the job, while keeping model identifiers and capabilities in configuration rather than evergreen campaign code.

Each generation becomes an immutable candidate record linked to the brief, prompt, model, model version where available, reference hashes, safety result, and generation time. The pipeline then performs format, duration, resolution, text, brand, duplicate, and policy checks. A reviewer sees the candidate with its lineage and approves a specific asset for a specific channel and schedule.

• Brief and claim authorization
• Configurable model and modality routing
• Queued candidate generation with bounded retries
• Immediate persistence of originals and provider metadata
• Technical, brand, policy, and accessibility validation
• Human approval tied to asset, channel, copy, audience, and schedule
• Authorized publication with idempotency and audit record
• Performance and rejection feedback returned to the brief and evaluation set

SOURCES: [2] OpenAI · [4] Google AI for Developers · [5] Google AI for Developers

Failure modes and model limits

OpenAI documents continuing limits in exact text placement, recurring character and brand consistency, and precise layout composition. Complex GPT Image requests may take up to two minutes; masks are guidance rather than exact geometry; transparent backgrounds are not supported; and output beyond 2560×1440 is experimental.

Google notes that image generation may not return the exact requested count and that real-person web-image grounding is unsupported. For Veo, audio processing can block generation, one request returns one output, and a seed does not create deterministic reproduction. These constraints favor candidate sets, explicit timeouts, asset inspection, and graceful partial completion.

SOURCES: [2] OpenAI · [4] Google AI for Developers · [5] Google AI for Developers

Security, rights, provenance, and publication control

Reference assets and uploaded material may contain confidential information, personal data, trademarks, or content without appropriate rights. Restrict source libraries, record authorization, avoid leaking sensitive assets into prompts or logs, and keep reviewer responsibility for claims, depictions, likeness, and context.

OpenAI states that its generated images include C2PA metadata and SynthID, while Google's current image and Veo documentation describes SynthID on generated media. OpenAI also warns that C2PA can be stripped and watermarks can degrade. Provenance is useful evidence of a generation path, but it does not establish factual truth, ownership, consent, or legal clearance.

SOURCES: [7] OpenAI · [8] OpenAI Help Center · [4] Google AI for Developers · [5] Google AI for Developers

Worked example: a week of campaign assets

A marketing owner approves three evidence-backed messages and a reference pack. The system creates still-image candidates using GPT Image 2 or Nano Banana 2 according to the asset requirements, then submits selected storyboards to Veo 3.1 Lite for short motion variants. Every result is copied to owned storage and attached to its brief and model record.

Automated checks reject wrong aspect ratios, missing mandatory text, prohibited claims, and unsafe outputs. A brand reviewer compares the asset to references, checks visible text and product representation, confirms rights and channel fit, and approves a specific post. Publication occurs through a scoped account connection. Rejections and performance become data for the next brief; they do not silently retrain or widen authority.

SOURCES: [2] OpenAI · [4] Google AI for Developers · [5] Google AI for Developers

Implementation checklist

Before enabling live publication, verify the following.

• Current model IDs, status, capabilities, limits, and deprecations live in configuration.
• The brief records claims, references, permissions, channel, and approval owner.
• Provider calls use bounded concurrency, timeouts, and retry rules that distinguish transient from user or moderation errors.
• Every generated asset is persisted immediately with lineage and immutable identifiers.
• Format, text, brand, accessibility, safety, and duplication checks run before review.
• A human approves the exact asset, copy, audience, channel, and schedule.
• Publication uses scoped account permissions, idempotency, and an audit record.
• Provenance is preserved where possible but is not treated as rights or truth verification.
• Preview models and deprecated APIs have feature flags and migration paths.

SOURCES: [2] OpenAI · [4] Google AI for Developers · [5] Google AI for Developers · [7] OpenAI

How Nautilus can help

Nautilus can design and implement the model router, media queue, storage and lineage, evaluation checks, approval experience, and authorized publishing integrations. We can also establish the brand and campaign operating rules, while leaving external publication and material claims under explicit client-controlled authority.