Gemini / Nano Banana (Google)

1. Introduction

1.1. Platform Overview

Nano Banana is Google DeepMind's product-name series for Gemini image models, spanning three tiers from entry-level to professional. Within the AI Harness six-layer capability model, Nano Banana is the strongest closed-source platform in this group on the L6 Governance & Security layer, and the only one with a publicly verifiable watermarking mechanism: all Google AI-generated images forcibly embed the invisible SynthID watermark, Nano Banana Pro supports the C2PA metadata standard, visible watermarks are tiered by subscription plan, and it restricts portraits of specific public figures under clear copyright protection.

It is also one of only two platforms in this group with quantified commercial case studies (the other is Meitu, with data from financial reports), which makes it a key sample for studying how "Harness capabilities translate into verifiable business results."

1.2. Model Lineage

Switching to the Thinking Model in the Gemini App uses Nano Banana Pro.

1.3. Open Access Forms and Pricing

Open access forms (consistent across channels, high confidence): Gemini App / Web, Google AI Studio, Gemini API, Vertex AI, Google Ads, Slides, Vids, NotebookLM, Flow, Google Search (AI Mode).

Subscription entry points (third-party sources, medium confidence): Google AI Plus $9.99/month (US), Google AI Pro $19.99/month; Ultra tier has higher quotas.

API pricing:

The official pricing page could not be scraped directly (ai.google.dev could not be fetched directly due to a login redirect) → official real-time pricing [To be filled]; must be re-verified against https://ai.google.dev/pricing before finalizing.

The table below contains values compiled from screenshots of the official pricing page as of 2026-02-26, for order-of-magnitude reference only and not to be used as contractual basis.

2. Glossary

2.1. Common AI Image Terms

2.2. Nano Banana-Specific Terms

2.3. Common Terms for Clothing Change and Face Swap

Note: Nano Banana does not provide a standalone virtual try-on or face-swap product capability, but its multi-reference consistency (14 reference images / 5 people) and search grounding make it one of the most commonly used generation foundations for third-party try-on solutions (such as THG Ingenuity AI Stylist, Max Fashion's digital fitting room, and ComfyUI's official VTON workflow).

3. Features

3.1. Generation and Editing Capabilities

3.2. Multi-Reference Consistency and Physical-Optics Control

• Multi-reference cap: fuses 14 reference images.
• Character consistency cap: maintains facial and clothing detail of up to 5 people in complex compositions. These two numbers are hard-coded context slots and represent the most explicitly publicized context-capacity contract in this group (for comparison: Kling O1 is 10 images, Runway Gen-4 is 3, FLUX.2 is 8~10, Midjourney is stackable but has no public cap).
• Physical and optics control: can specify lighting atmosphere, lens type, depth of field, viewing angle, and composition ratio - turning photography parameters into first-class prompt fields.

3.3. Structured Production Brief

Officially, it is recommended to write prompts as a structured production brief, with fields including:

This is conceptually consistent with FLUX.2's official JSON fields (subject / background / lighting / style / camera_angle / composition): upgrading context from free-form text into a verifiable field contract. The difference is that FLUX.2 is machine-readable JSON, while Nano Banana is the officially recommended natural-language authoring convention.

3.4. Known Limitations

1. Cannot generate content involving violence or hate incitement.
2. Cannot generate portraits of specific public figures under clear copyright protection - an identity restriction rarely seen in this group and explicitly stated by the vendor.
3. Visible watermark retained by tier: the Gemini starburst watermark is retained for the free tier and Google AI Pro; removable for Ultra subscribers or images generated via AI Studio / Vertex AI.
4. Pricing must be re-verified in real time: the official pricing page could not be scraped directly due to a login redirect, [To be filled].

4. Platform Architecture

图 4-1｜Nano Banana 平台架构：入口—推理—底座三层

数据来源：基于本文分析绘制的示意图。

4.1. Native Multimodal Foundation

• The foundation is the Gemini 3 Pro / 3.1 Flash multimodal LLM, natively multimodal (mixed image-text input and output), rather than an attached diffusion pipeline.
• This difference has cascading effects:
• No CFG / Steps / Sampler diffusion-paradigm parameters;
• No LoRA / ControlNet / IP-Adapter ecosystem (these are supporting mechanisms of diffusion models);
• Consistency relies on the model's own contextual understanding, not external condition injection.

4.2. Reasoning Paradigm

By default it includes reasoning steps: first understand intent and image logic, retrieve from the web when necessary (Grounding), then generate. This is the same paradigm as Jimeng Seedream 5.0's "web retrieval + CoT chain-of-thought reasoning", and is the third-generation generation-architecture consensus visible in this group - think first, then draw.

4.3. Distribution Matrix

Key enterprise-side difference: images generated via Vertex AI can have the visible watermark removed, and support enterprise security mechanisms and elastic billing. This means the same model has different compliance and governance attributes under the "consumer entry point" versus the "enterprise entry point" - a key point that is easy to overlook during selection.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Nano Banana's L1 has three measurable characteristics, which are extremely rare in this group:

1. An explicit capacity contract: 14 reference images, 5 people. This is not a vague statement like "supports multiple images", but a hard cap that can be written into technical specs and acceptance criteria.
2. Ultra-long prompts: supports 64k input tokens, letting you stuff an entire brand guideline, product catalog, or historical-asset description into context at once.
3. External knowledge grounding: Search Grounding brings Google search results into the generation context, making "generating images from real-time facts" possible (weather, events, recipes, specialized knowledge bases).

The officially recommended structured production brief (9 fields) further turns context organization from a "writing skill" into a "filling spec". Apart from FLUX.2's official JSON contract, this is the only context-organization standard given by a platform vendor in this group.

5.3. L2 Tools & Execution Layer

• Toolset: generation, editing, retrieval (Grounding), multilingual rendering.
• Exposure: Gemini API and Vertex AI. The Vertex AI side additionally provides enterprise security mechanisms and elastic billing.
• No open tool registration: external tools cannot be registered into the generation flow (by contrast, Leonardo.ai supports registering custom fine-tuned models as callable model IDs).
• Ecosystem integration: aggregated by third-party platforms and workflows (e.g., ComfyUI's official VTON workflow uses Nano Banana Pro as a generation node), showing that its API contract has good composability.

5.4. L3 Orchestration & Control Layer

• Multi-turn conversational iteration: generate -> critique with text -> targeted revision.
• Explicit vendor operational discipline: "request only one controlled change at a time". This is a rare official orchestration guideline - it raises "change granularity" to a usage norm, effectively managing context-drift risk: if you change several things at once, uncontrolled parts may also be changed by the model along the way.
• Still missing: no DAG, no sub-agent dispatch, no workflow-artifact export and version control, no interruption recovery.

Compared with Kling's "generation as conversation", the two share the same paradigm; the difference is that Kling emphasizes "no mask needed", while Nano Banana emphasizes "one change at a time".

5.5. L4 Memory & State Layer

• Relies on Google ecosystem persistence: Drive, NotebookLM, Workspace.
• No first-class abstraction for "brand asset packs" - an obvious shortcoming compared with Runway's Brand Kits (up to 3), Kling's subject creation (metered pricing), and Meitu Design Studio's creative-asset reuse (explicitly stated).
• No public documentation of Checkpoint, artifact version management, or a model-version locking endpoint.

Engineering implication: if a business needs assets such as a "locked brand character / locked logo / locked palette", Nano Banana has no native abstraction, so the business system must implement it itself and re-inject context on every request - which increases 64k-token context usage and operational cost.

5.6. L5 Evaluation & Observability Layer

• The vendor emphasizes manual verification steps: checking every word, number, face and product detail before release. This is effectively a vendor-issued acceptance checklist; although it is not an automated Eval Set, it is actionable and executable, better than the "no evaluation mechanism" status of most platforms.
• No public official Eval Set, Golden Dataset, or regression set.
• Observability basis: on the API side, pricing is per resolution and token, so cost is measurable; but the official pricing page could not be scraped directly, [To be filled].

5.7. L6 Governance & Security Layer

This is the layer where Nano Banana is decisively ahead in this group, and it is the only closed-source platform here that can correspond to both the explicit and implicit dual-labeling requirements of China's Labeling Measures:

Must be stated honestly: the mapping of the above mechanisms to the Labeling Measures is a comparative analysis made by this report based on public materials, not a compliance statement by Google regarding Chinese regulations. Whether Google has implemented specific metadata fields such as "service provider name or code, content ID" as required by China's Labeling Measures has no public statement, [To be filled].

At the same time, the visible watermark is removable at the Ultra / AI Studio / Vertex AI entry points, which is a reasonable design on the enterprise side (enterprise-owned assets), but it also means that relying only on platform default behavior is insufficient to satisfy the explicit-labeling obligation, and the business side still needs to build its own injection step.

5.8. Maturity Assessment

Nano Banana takes the form of "strong model + upper-middle Harness + strongest governance". Its L1 has a clear capacity contract and grounding capability, L3 has official operational discipline, L5 has an executable official acceptance checklist, and L6 is the only closed-source platform in this group with the triple mechanism of "mandatory invisible watermark + metadata standard + verifiable entry point". The shortcoming is L4 - the lack of a first-class brand-asset abstraction, which requires an additional external hosting layer in production scenarios that need "long-term, multi-batch, asset-locking".

6. Practical Cases

6.1. THG Ingenuity × Google Cloud "AI Stylist"

• Solution: a virtual try-on solution built on the Gemini Enterprise Agent Platform, already listed on the Google Cloud Marketplace.
• Quantified data (early Myprotein deployment): for users who used AI Stylist at least once, compared with the in-site average for sportswear purchases -
• UK users are close to 6× more likely to convert;
• On-site dwell time is 6.3×;
• Average order value is 2.5% higher.
• Source: Prolific North, high relevance (contains quantified data).
• Nature: the data comes from public reports on the solution provider's early deployment and is vendor-disclosed data; the source and basis should be noted when citing.

6.2. Max Fashion × Google Cloud

• Solution: Middle East retail giant Max Fashion launched an enterprise-grade digital fitting room built on the Gemini Enterprise platform, generating highly realistic real-time previews that show how garments drape and move across diverse, real body types.
• Quote (Bala Subramaniam, Senior Vice President of Omnichannel at Max Fashion): "Virtual try-on fundamentally closes the distance between online and offline experiences. Customers browsing on their phones can, for the first time, have the same shopping confidence as if standing in a physical fitting room."
• Source: fabrix.hk, citing Retail Management Middle East, medium-to-high relevance.
• Quantified data: no specific conversion or return-rate figures disclosed.

6.3. L'agence × Google AI (February 2026 New York Fashion Week)

• Content: guests can instantly visualize themselves wearing the runway looks, with instant booking supported.
• Source: Business of Fashion, medium-to-high relevance.
• Quantified data: not disclosed.

7. Summary

7.1. Strengths

1. L6 governance is the strongest in the group: SynthID's mandatory invisible watermark (officially stated as non-removable) + C2PA + tiered visible watermark + public-figure portrait restrictions + an in-app verifiable entry point, forming a complete "generation—labeling—verifiable" closed loop.
2. The context capacity contract is explicit: 14 reference images / 5 people / 64k tokens, which can be written into technical specifications and acceptance criteria.
3. Search Grounding: brings real-time external facts into the generation context, supporting infographics and data visualization.
4. The official acceptance checklist is usable: word-by-word, number-by-number, face-by-face, and product-detail-by-product-detail verification, directly convertible into an enterprise DoD.
5. It is one of only two platforms with quantified commercial case studies: THG Ingenuity's conversion is close to 6× and dwell time 6.3×.
6. The widest distribution matrix: from consumer-side Search / Workspace to developer-side API / Vertex AI, and the enterprise entry point (Vertex AI) is independent of the consumer entry point in compliance and billing.

7.2. Limitations and Applicability Boundaries

Applicability boundaries: suitable for scenarios that require strong compliance backing, real-time fact grounding, and consumer-level reach (Search / Workspace); not suitable for production scenarios that require long-term locking of brand assets, fully offline or privatized deployment, or precise structured control (such as FLUX.2's hex brand colors).

7.3. Selection Recommendations

1. Always use Vertex AI on the enterprise side: the visible watermark is removable, enterprise security mechanisms and elastic billing are supported, and the consumer entry point differs in compliance attributes.
2. Build your own brand-asset layer: consolidate brand guidelines, character references, and color palettes into an external asset library, and inject them on every request in the official 9-field production-brief format.
3. Follow the "one change at a time" discipline: in multi-round iteration, request only one controlled change at a time, avoiding uncontrolled regions being modified along the way.
4. Implement the acceptance checklist: write the official "word-by-word, number-by-number, face-by-face, and product-detail-by-product-detail verification" into the DoD, especially when text rendering and faces are involved.
5. Pricing must be re-verified in real time: before finalizing the document and before procurement, visit the ai.google.dev pricing page to confirm current unit prices.
6. Do not rely on the platform's default labeling: even with SynthID forcibly embedded, the explicit labeling and specific metadata fields required by China's Labeling Measures must still be injected by the business side itself in the export chain.

7.4. Compliance Notes

• The "Measures for Labeling AI-Generated Synthetic Content" (CAC General Office Document No. 2 of 2025): Article 4, the explicit-labeling obligation; Article 5, the implicit-labeling obligation (including the service provider's name or code and the content number); Article 6, the distribution platforms' verification obligation; Article 10, prohibition on maliciously deleting, altering, forging, or concealing labels.
• Article 1019 of the Civil Code of the People's Republic of China: no one may infringe another person's portrait right by forging or other means using information technology. Nano Banana's restriction on portraits of public figures is a platform-side guardrail and cannot substitute for the rights holder's authorization.
• The 2026-03 effective judgment of the Beijing Internet Court: identifiability + burden-of-proof shift. Even if the platform has public-figure restrictions, enterprises should still retain generation parameters, reference images, and prompts in preparation for evidentiary proceedings.

Information Gap Statement

1. Official pricing page: ai.google.dev could not be fetched directly due to a login redirect; official real-time pricing [To be filled]; the values in the table are compiled from a 2026-02-26 screenshot.
2. C2PA support: per the Baidu Baike statement, not the original text of Google's official documentation.
3. Availability of each model in China: no official statement found on the accessibility and compliance path of the Gemini App / API in mainland China. [To be filled]
4. Compliance mapping to the Labeling Measures: a comparative analysis made by the report based on public materials, not an official Google compliance statement; whether specific metadata fields such as "service provider name or code, content ID" are implemented [To be filled].
5. Seed locking, model-version locking endpoint, snapshot mechanism: no official statement found. [To be filled]
6. Quota details for Google AI Plus / Pro / Ultra: only third-party sources give prices; quotas are unspecified. [To be filled]
7. Release date of Nano Banana 2 (Gemini 3.1 Flash Image): search results only note "new tier" without a date. [To be filled]

8. References

1. Google AI · Gemini API pricing page (official, must be re-verified before finalizing) — Google. https://ai.google.dev/pricing
2. The Rundown AI · Nano Banana Pro capabilities, pricing and entry points (third-party, medium confidence) — https://www.therundown.ai/tools/nano-banana-pro
3. iKala · Nano Banana Pro features and solution introduction (incl. model comparison table, third-party) — https://ikala.ai/zh-tw/blog/ikala-ai-insight/nano-banana-pro-functions-how-to-use/
4. Brand Icon Image · Google Introduces Nano Banana Pro (official launch-event retelling, third-party) — http://www.brandiconimage.com/2025/11/google-introduces-nano-banana-pro-with.html
5. FoneArena · Google rolls out Nano Banana Pro powered by Gemini 3 (third-party) — https://www.fonearena.com/blog?p=469327
6. Baidu Baike · Nano Banana Pro (incl. SynthID / C2PA / 14 reference images / 5 people, secondary source) — https://baike.baidu.com/item/BananaPro/67405719
7. Prolific North · "THG Ingenuity predicts 600% conversion rate boost with Google AI Stylist rollout" (incl. quantified data) — https://www.prolificnorth.co.uk/news/thg-ingenuity-predicts-600-conversion-rate-boost-with-google-ai-stylist-rollout/
8. Business of Fashion · "How To Drive Luxury Conversion Through AI-Powered Virtual Try-On" (incl. L'agence case) — https://www.businessoffashion.com/articles/technology/ai-virtual-try-on-luxury-dressx-report/
9. FabriX · "How AI Is Solving Fashion's Fitting Room Crisis & Driving E-Commerce Sales" (incl. Max Fashion × Google Cloud) — https://fabrix.hk/blog/how-ai-is-solving-fashions-fitting-room-crisis
10. ComfyUI official workflow 「虚拟角色试穿 - 四合一」 (uses Nano Banana Pro as the generation node) — Comfy Org. https://comfy.org/zh/workflows/templates_rob_fashion_shoot_vton-4in1.app/
11. "Measures for Labeling AI-Generated Synthetic Content" (CAC General Office Document No. 2 of 2025) — Cyberspace Administration of China et al., effective 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
12. Article 1019 of the Civil Code of the People's Republic of China — National People's Congress, 2020. (article cited in body text, no official link)
13. Economic Information Daily · "Technology is not an infringement 'shield'; the court finds AI 'face theft' this way" — Xinhua News Agency Economic Information Daily, 2026-04-17. http://dz.jjckb.cn/www/pages/webpage2009/html/2026-04/17/content_115180.htm