Seedream 5.5 is live on Atlas Cloud — native transparent PNG & 80+ round editing
Home
Explore
Google
Nano Banana 2.1
google/nano-banana-2.1-developer/reference-to-image
Nano Banana 2.1 Developer Reference-to-Image
image-to-image
DEV

Nano Banana 2.1 Developer Reference-to-Image API by Google

google/nano-banana-2.1-developer/reference-to-image
Reference-to-image

Google's Nano Banana 2.1, built on Gemini 3.6 Flash, turns one source video clip, optional reference images and a prompt into a new still image such as a thumbnail, poster or summary infographic, with Pro-level quality at Flash-level speed.

Nano Banana 2.1 Developer Reference-to-Image is developed by Google. Atlas Cloud (operated by Atlas Cloud AI LLC) provides access to it and does not own it. All trademarks belong to their respective owners.

Nano Banana 2 — Pro Quality at Flash Speed

V2

Also known as Gemini 3.1 Flash Image

Google DeepMind's latest image model combines the advanced capabilities of Nano Banana Pro with the speed of Gemini Flash — delivering 3-5x faster generation, up to 4K resolution, and character consistency for up to 5 characters in a single workflow.

Next-Gen Image Generation
  • Up to 4K resolution output (512px / 1K / 2K / 4K tiers)
  • 10+ aspect ratios including 21:9, 1:4, 8:1 and more
  • Accurate and legible text rendering within images
  • Near Pro-level quality (~95%) at Flash speed
Intelligent Editing & Consistency
  • Character consistency for up to 5 characters across scenes
  • Object fidelity for up to 14 objects in one workflow
  • Targeted edits via natural language (remove, replace, repose)
  • Multi-image blending and seamless composition

What's New in Nano Banana 2

3-5x Faster Than Pro

Built on the Gemini 3.1 Flash architecture, Nano Banana 2 generates standard images in 4-8 seconds — compared to 10-20 seconds for Nano Banana Pro.

Image Search Grounding

The standout feature of NB2 — it can retrieve real-world reference images via Google Search during generation, significantly improving accuracy for landmarks, famous people, and brand logos.

Accurate Text Rendering

Generate accurate, legible text for marketing mockups, greeting cards, and localized content. You can even translate and localize text within an image.

Multi-Character Consistency

Maintain visual consistency for up to 5 characters and 14 objects across scenes — perfect for storyboards, comics, and marketing campaigns.

Prompt Examples & Templates

Explore curated prompt templates showcasing Nano Banana 2's key capabilities — text rendering, character consistency, search grounding, and 4K output.

Marketing Mockup with Text
Text Rendering

Marketing Mockup with Text

Generate marketing visuals with accurate, legible text — one of NB2's standout improvements
Prompt

A minimalist coffee shop promotional poster with the text 'MORNING BREW — Fresh Roasted Daily' in elegant serif font, warm earth tones, steam rising from a ceramic cup, clean layout with plenty of whitespace

Multi-Scene Character
Character Consistency

Multi-Scene Character

Maintain character consistency across multiple scenes — supports up to 5 characters per workflow
Prompt

A young woman with short red hair and freckles, wearing a green jacket, standing in a rainy Tokyo street at night with neon reflections on wet pavement, cinematic lighting, photorealistic

Person to Action Figure
Photo to Action Figure

Person to Action Figure

Transform people from photos into collectible action figures with custom packaging
Prompt

Transform the person in the photo into an action figure, styled after [CHARACTER_NAME] from [SOURCE / CONTEXT]. Next to the figure, display the accessories including [ITEM_1], [ITEM_2], and [ITEM_3]. On the top of the toy box, write "[BOX_LABEL_TOP]", and underneath it, "[BOX_LABEL_BOTTOM]". Place the box in a [BACKGROUND_SETTING] environment.

Real-World Reference Generation
Search Grounding

Real-World Reference Generation

Leverage Image Search Grounding to generate accurate real-world subjects like landmarks and brands
Prompt

A photorealistic aerial view of the Eiffel Tower at golden hour, with the Seine River winding through Paris below, warm sunset light casting long shadows, high detail, 4K resolution

Product Design Render
Product Photography

Product Design Render

Create professional product photography with precise control over lighting and composition
Prompt

A frosted glass perfume bottle with a marble cap on a white marble surface, soft studio lighting from the left, subtle reflections, minimalist luxury aesthetic, product photography style

Artistic Style Transformation
Style Transfer

Artistic Style Transformation

Apply diverse artistic styles while maintaining subject integrity
Prompt

Transform this photo into Studio Ghibli animation style, keeping the same composition and subjects, lush watercolor backgrounds, soft diffused lighting, whimsical atmosphere

Ultra High Resolution Scene
4K Output

Ultra High Resolution Scene

Generate detailed scenes at up to 4K resolution with rich textures
Prompt

A cozy Japanese ramen shop interior at night, steam rising from bowls, warm amber lighting, detailed wooden counter with various condiments, a chef working in the background, 4K, ultra detailed

Use Cases

🎬
Storyboarding & Comics
📸
Product Photography
📊
Marketing Mockups
📱
Social Media Content
🔤
Text Overlay Design
👤
Character Design
✨
Photo Editing & Retouching
🎨
Brand Visual Content

Why Choose Nano Banana 2?

⚡

Flash Speed

3-5x faster than Nano Banana Pro with 4-8 seconds standard generation time
🎯

Near-Pro Quality

Achieves approximately 95% of Pro's image quality in most scenarios
💰

Cost Effective

Roughly half the cost of Nano Banana Pro — making high-quality AI image generation more accessible

Technical Specifications

Architecture:Gemini 3.1 Flash (GEMPIX2)
Resolution Support:512px to 4K (512px / 1K / 2K / 4K tiers)
Aspect Ratios:1:1, 4:3, 3:4, 2:3, 3:2, 16:9, 9:16, 1:4, 4:1, 8:1, 21:9
Consistency:Up to 5 characters + 14 objects per workflow
Content Safety:SynthID watermark, C2PA standard compatible
API Access:Gemini API, Vertex AI, AI Studio, Gemini CLI

Experience Nano Banana 2

Pro-level image generation at Flash speed — create stunning visuals with character consistency, text rendering, and 4K resolution support.

✨Free Credits to Start
⚡Instant API Access
🌐No Setup Required

Google Nano Banana 2.1 Developer — Reference-to-Image

Nano Banana 2.1 (gemini-nano-banana-2.1) is Google DeepMind's latest image generation and editing model, released and made generally available on October 6, 2026 as the successor to Nano Banana 2 (Gemini 3.1 Flash Image). Built on Gemini 3.6 Flash and part of the Gemini 3 model family, it delivers Pro-level image generation and editing at Flash-level speed. It is a multimodal reasoning image model: it understands text, images and video, and thinks through complex spatial and compositional instructions before it generates an image. Compared with Nano Banana 2, Google reports better instruction following, generation consistency, multi-reference composition and visual factuality. In Google's side-by-side human evaluations it outperforms both Nano Banana 2 and Nano Banana Pro across every published text-to-image and editing benchmark, with the biggest gains in visual design, mask-based editing, subject consistency and infographic factuality.

This Reference-to-Image endpoint is a video-to-image workflow. It drives the model with one source video clip, optional reference images and a natural-language prompt. The model samples frames from the clip, works out its subjects, scenes and key moments, and generates a new still image from them, such as a thumbnail, a cinematic poster, a summary infographic or new artwork inspired by a scene. It runs the same model as the Text-to-Image and Edit endpoints; only the inputs differ.

This page covers the Developer edition of the endpoint (google/nano-banana-2.1-developer/reference-to-image). It runs the same model, accepts exactly the same parameters, and uses the same unit prices and pricing formula as the standard Nano Banana 2.1 — Reference-to-Image endpoint.

🌟 What's New in Nano Banana 2.1

  • Video as a visual reference: Pass a public YouTube URL or an HTTP video URL (up to 25 minutes long), trim it to the segment that matters, and set the frame sampling rate.
  • Subject consistency: Add up to 14 reference images alongside the video. The model keeps up to 4 characters and 10 objects consistent, so the people and products in your footage stay recognizable.
  • Better visual design: Posters, thumbnails, Swiss-style grid layouts and infographics come out as complete, well-composed designs instead of isolated elements.
  • More natural-looking images: Better physical plausibility in scene construction and higher overall image quality.
  • Precise, multilingual text rendering: Legible titles, captions and labels in the generated image, with localized text rendering across several languages.
  • Real-world knowledge and search grounding: The model draws on Gemini's world knowledge and can ground the output with real-time Google Web Search and Google Image Search. Grounding improves recognition of long-tail entities, accuracy and visual fidelity.
  • Flexible formats: 1k, 2k and 4k output in 14 fixed aspect ratios plus auto, including ultra-wide and ultra-tall formats (21:9, 4:1, 1:4, 8:1, 1:8).
  • Adjustable thinking: The model thinks through complex spatial and compositional instructions before it generates an image. Choose how much it reasons before it draws. Use minimal for speed and high for summary infographics and complex compositions.

📊 Benchmark Highlights

The results below come from Google DeepMind's Nano Banana 2.1 model card (October 2026). Elo scores come from side-by-side human evaluations, and Infographic Factuality is scored by a single-sided autorater. In the tables, "Nano Banana 2" is Gemini 3.1 Flash Image (with thinking) and "Nano Banana Pro" is Gemini 3 Pro Image.

Text-to-Image

BenchmarkNano Banana 2.1 (Thinking)Nano Banana 2.1 (No Thinking)Nano Banana 2Nano Banana Pro
Overall Preference (Elo)1050 ± 141015 ± 13990 ± 7935 ± 8
Infographic Design (Elo)1048 ± 171001 ± 17961 ± 12912 ± 12
Infographic Factuality0.5210.3280.1790.265

Editing

BenchmarkNano Banana 2.1 (Thinking)Nano Banana 2.1 (No Thinking)Nano Banana 2Nano Banana Pro
General Editing (Elo)1026 ± 12980 ± 15938 ± 11939 ± 10
Single-Character Consistency (Elo)1028 ± 141021 ± 14981 ± 10991 ± 9
Multi-Character Consistency (Elo)1106 ± 141068 ± 14978 ± 101011 ± 10
Mask/Ink-Based Editing (Elo)1049 ± 151042 ± 16965 ± 12927 ± 12
Product Consistency (Elo)1024 ± 18981 ± 18955 ± 22965 ± 14
Stylization (Elo)1062 ± 201036 ± 17991 ± 12990 ± 12
Multi-Reference Editing (Elo)1066 ± 221041 ± 20988 ± 13989 ± 12

In every category the thinking configuration scores higher than the no-thinking one.

📐 Technical Specifications

The table below summarizes Google's published specifications for the underlying model.

ItemSpecification
Model IDgemini-nano-banana-2.1
Base modelGemini 3.6 Flash (Gemini 3 series)
Launch stageGenerally available (GA) since October 6, 2026
Input modalitiesText, image, video and PDF (audio is not supported)
Output modalityImage
Token limits128K input tokens (text and images), 32K output tokens
Image inputUp to 14 images per request (bounded by the input token limit), in PNG, JPEG, WebP, HEIC or HEIF format, up to 7 MB each
Video inputHTTP or YouTube URL, up to about 25 minutes long, in MP4, MOV, MPEG, WebM, FLV, WMV or 3GPP format
Output resolutions1K (about 1 megapixel), 2K (about 4 megapixels), 4K (about 16 megapixels)
Aspect ratios1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 4:1, 1:4, 8:1, 1:8 (supported range: 1:8 to 8:1)
ThinkingSupported, with minimal, medium (default) and high levels
GroundingGoogle Search (web) and Google Image Search
Image capabilitiesText-to-image, image editing, image generation from video input, virtual try-on, person generation
Content credentialsSynthID watermark and C2PA, applied automatically
Not supportedAudio input, transparent (alpha channel) output
Knowledge cutoffMarch 2026

This endpoint takes a text prompt plus exactly one video clip, with up to 14 optional reference images.

🎬 How It Works

The model reads the part of the clip between start and ends, sampling frames at the rate set by fps. Each sampled frame is processed at the detail level chosen by media_resolution. Together with your prompt and any reference images, these frames form the model's multimodal context. The model then generates a single new image grounded in the video's subjects, style and key moments. Lower fps and media_resolution values let longer footage fit in context, and higher values capture fast action and fine detail.

⚙️ How to Use

  1. Add the video clip. Provide exactly one item in video_clips with its url, a start and ends time in seconds (ends = 0 uses the whole video), and a sampling fps.
  2. Write your prompt. Describe the image you want from the footage, for example a thumbnail, poster, infographic or artwork. Put any exact text to render in quotes.
  3. Add reference images (optional). Provide 1–14 images in reference_images for characters, logos, products or styles that must look exact.
  4. Pick a format. Choose an aspect_ratio and a resolution (1k, 2k or 4k).
  5. Fine-tune (optional). Set thinking_level, enable web or image search, or adjust media_resolution.
  6. Run. The result is returned as an image URL, or as a BASE64 string through the API when enable_base64_output is set.

Example Request Body

{ "model": "google/nano-banana-2.1-developer/reference-to-image", "prompt": "Create a 16:9 video thumbnail from the most exciting moment of this clip, with the bold title \"FINAL LAP\" in the upper left", "video_clips": [ { "url": "https://www.youtube.com/watch?v=VIDEO_ID", "start": 30, "ends": 90, "fps": 1 } ], "aspect_ratio": "16:9", "resolution": "2k", "thinking_level": "high" }

🧩 Parameters

Core Inputs

ParameterTypeRequiredDefaultAllowed Values / RangeDescription
modelstringYesgoogle/nano-banana-2.1-developer/reference-to-image—model name
promptstringYes——The prompt describing the image to generate from the reference video clip and optional reference images, e.g., a thumbnail, cinematic poster, summary infographic or new artwork inspired by a video scene.
video_clipsarray of objectYes—Exactly 1 itemSource video clips to use as references for generation. Supports 1 video clip.
reference_imagesarray of stringNo—1 – 14 itemsOptional images to use as character, object or style references alongside the video. Accepts 1 to 14 images; the model keeps up to 4 characters and 10 objects consistent. Supported formats: PNG, JPEG, WebP, HEIC, HEIF. Each image is limited to 7MB. Supports both a public URL and a base64-encoded image for each item.

Video Clip Fields

Each item in video_clips is an object with the following fields, all of which are required:

FieldTypeRequiredDefaultAllowed Values / RangeDescription
urlstring (uri)Yes——URL of the source video clip. Support HTTP URL or YouTube video URL. The video length is limited to 25 minutes.
startnumberYes0≥ 0Start time in seconds for trimming the video clip.
endsnumberYes0≥ 0End time in seconds for trimming the video clip. Set 0 to clip the whole video
fpsnumberYes10 – 24FPS of the video clip.

Output Settings

ParameterTypeRequiredDefaultAllowed Values / RangeDescription
aspect_ratiostringNoautoauto, 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9, 4:1, 1:4, 8:1, 1:8The aspect ratio of the generated image. "auto" lets the model decide: it matches the input image when editing, otherwise it generates a 1:1 image. Ultra-wide and ultra-tall ratios (21:9, 4:1, 1:4, 8:1, 1:8) are supported.
resolutionstringNo1k1k, 2k, 4kThe resolution of the output image: 1k (about 1 megapixel), 2k (about 4 megapixels) or 4k (about 16 megapixels).

Reasoning & Grounding

ParameterTypeRequiredDefaultAllowed Values / RangeDescription
thinking_levelstringNodefaultdefault, minimal, medium, highControls the amount of internal reasoning the model performs before generating the image. Higher levels improve quality on complex prompts (e.g., infographics, multi-character scenes, dense text) but increase latency. "default" uses the model's default level (medium); "minimal" is the fastest.
enable_web_searchbooleanNofalsetrue, falseIf enabled, the model will use Google web search to ground the generation with real-time information (e.g., current events, weather, stock charts).
enable_image_searchbooleanNofalsetrue, falseIf enabled, the model will use images retrieved via Google Image Search as visual context for the generation. Can be used alone or together with web search. Real-world images of people from search results are not used.

Advanced Settings

ParameterTypeRequiredDefaultAllowed Values / RangeDescription
media_resolutionstringNodefaultdefault, low, medium, highControls how input images and videos are processed. "low" uses fewer tokens per image / video frame, possibly losing detail but allowing longer videos in context; "high" preserves more detail. "default" lets the model decide.
top_pnumberNo0.950 – 1Probability threshold for top-p sampling
temperaturenumberNo10 – 1Creativity allowed in the responses. Best results at default 1.0. Lower values may impact reasoning.
seedintegerNo-1Any integerRandom seed. Does not guarantee determinism but may improve repeatability. -1 means a random seed will be used.

API-Only Options

ParameterTypeRequiredDefaultAllowed Values / RangeDescription
enable_sync_modebooleanNofalsetrue, falseIf set to true, the function will wait for the result to be generated and uploaded before returning the response. It allows you to get the result directly in the response. This property is only available through the API.
enable_base64_outputbooleanNofalsetrue, falseIf enabled, the output will be encoded into a BASE64 string instead of a URL. This property is only available through the API.

💰 Pricing

Each request is billed for the output image at its resolution, plus a fee for the source video clip, a small fee for each optional reference image and a flat fee for each search grounding option you enable. Aspect ratio, thinking level, media resolution, sampling settings and seed do not affect the price.

Unit Prices

SKUDescriptionUnit Price
sku_1k1k output image$0.05 per image
sku_2k2k output image$0.075 per image
sku_4k4k output image$0.14 per image
sku_video_clipSource video clip in video_clips$0.20 per clip
sku_input_imageReference image in reference_images$0.00168 per image
sku_web_searchGoogle Web Search grounding$0.014 per request, when enabled
sku_image_searchGoogle Image Search grounding$0.014 per request, when enabled

Pricing Formula

cost = (resolution == "2k" ? sku_2k : (resolution == "4k" ? sku_4k : sku_1k)) * num_images + (enable_web_search ? sku_web_search : 0) + (enable_image_search ? sku_image_search : 0) + sku_video_clip * len(video_clips) + sku_input_image * len(reference_images)
  • Resolution charge: resolution selects the per-image price: 2k costs 0.075,‘4k‘costs0.075, `4k` costs 0.14, and 1k (the default) costs $0.05.
  • Number of images: num_images is not an input on this endpoint, so it keeps its default value of 1. Each request is billed for exactly one output image.
  • Video clip: Each clip in video_clips costs 0.20.Exactlyoneclipisrequired,soeveryrequestincludesa0.20. Exactly one clip is required, so every request includes a 0.20 video charge. The charge does not depend on the clip's length, the start / ends trim range or the fps.
  • Reference images: Each image in reference_images adds 0.00168.Thefieldisoptional,soitaddsnothingwhenomittedandupto0.00168. The field is optional, so it adds nothing when omitted and up to 0.02352 with 14 images.
  • Web search: A flat $0.014 is added once per request when enable_web_search is true.
  • Image search: A flat $0.014 is added once per request when enable_image_search is true. You can enable both search options together.

Examples

ResolutionVideo ClipsReference ImagesWeb SearchImage SearchCalculationTotal
1k10OffOff0.05+0.05 + 0.20$0.25
2k10OffOff0.075+0.075 + 0.20$0.275
4k10OffOff0.14+0.14 + 0.20$0.34
1k13OffOff0.05+0.05 + 0.20 + 3 × $0.00168$0.25504
2k12OnOff0.075+0.075 + 0.20 + 2 × 0.00168+0.00168 + 0.014$0.29236
4k114OnOn0.14+0.14 + 0.20 + 14 × 0.00168+0.00168 + 0.014 + $0.014$0.39152

💡 Best Use Cases

  • Video thumbnails: Generate eye-catching, on-brand thumbnails from the footage itself, with legible titles.
  • Cinematic posters and key art: Turn a standout moment into a polished poster or promotional still.
  • Summary infographics: Distill a tutorial, lecture or product demo into a single explanatory image.
  • Social media assets: Produce square, vertical and landscape stills from one video for every platform.
  • Brand and campaign content: Keep characters, products and logos consistent with your video campaigns by adding them as reference images.
  • Concept art and storyboards: Derive new artwork inspired by a specific scene while keeping its characters and mood.

✍️ Prompting Tips

  • Trim to the moment that matters. Set start and ends around the key scene so the model focuses on it. Set ends to 0 only when you need the whole video.
  • Match fps to the footage. A low fps, such as the default 1, works well for long or slow clips. Raise it for short, fast-action segments.
  • Balance detail and length. Use media_resolution: low for long videos and high when fine detail in the frames matters.
  • Name the deliverable. Be specific about the output, for example "a 16:9 YouTube thumbnail", "a 2:3 movie poster" or "a 9:16 story cover".
  • Describe what you want, not what you don't. Phrase the scene positively: instead of "no cars", write "an empty, deserted street".
  • Quote exact text. Wrap titles and captions in quotes, and say where they go and how they look.
  • Anchor identity with reference images. Add a clean image of the main character, product or logo when it has to match exactly.
  • Use high thinking for complex outputs. high helps summary infographics and multi-character compositions.

🔒 Content Authenticity

Every image created or modified with Nano Banana 2.1 carries an imperceptible SynthID watermark and is automatically signed by Google LLC with C2PA Content Credentials, which list "Google Media Processing Services" as the creating app. Both identify the image as AI-generated: SynthID is embedded directly in the pixels, while the C2PA metadata is cryptographically signed, tamper-evident and can be inspected in any C2PA-compatible app. Modifying the file with a tool that doesn't support C2PA invalidates the credentials.

⚠️ Limitations

  • Footage quality: Fast motion, low-quality footage or a very low sampling rate can produce less precise results, because the model only sees the sampled frames.
  • Character consistency: Consistency between the input video, the reference images and the output is strong but not always perfect.
  • Spatial reasoning: The model occasionally confuses left and right, and it can get the scale of a subject wrong relative to its surroundings.
  • Small and long text: Small text (often blurry at 1k), long paragraphs and page-length text can still render poorly.
  • World knowledge: Advanced world knowledge, 3D reasoning and factuality are still limited. The underlying knowledge cutoff is March 2026, and some domains may only reflect knowledge up to January 2025.
  • No transparent backgrounds: The model can't output images with an alpha channel, so it can't produce true transparent PNGs, for example for logos or stickers.
  • Latency: Requests can occasionally be slow or time out, especially with long clips, high fps or high thinking levels.

📝 Notes

  • prompt and video_clips are both required. video_clips must contain exactly one clip.
  • The video url can be an HTTP URL or a YouTube URL, and the video can be up to 25 minutes long.
  • reference_images is optional. When provided, it accepts 1–14 items in PNG, JPEG, WebP, HEIC or HEIF format. Each image can be up to 7MB and can be a public URL or a base64-encoded image.
  • Each request returns a single image.
  • Prompts, video clips and reference images must comply with Google's Safety Guidelines. If a request fails, review your inputs for restricted content, adjust them and try again.

Explore Similar Models

One API for All Media AI.

Explore all models