Is Nano Banana Video Real? Workflow, Cost and Code (2026)

A nano banana video is a short AI clip whose starting frame, or every keyframe, was made with Google’s Nano Banana image model and then animated by a separate video model. That second part surprises a lot of people. Nano Banana is the nickname for Google’s Gemini image models, and none of them output motion. If you searched for this term, you have probably landed on sites that sell a “Nano Banana video generator”, and it helps to know what is actually happening behind them before you pay for anything.

This guide explains what the term really means, which Nano Banana model to use for video work, how the keyframe-to-clip workflow goes, what it costs per frame, and how to script the image side through an API.

What is a nano banana video?

“Nano Banana” started as a codename. Google’s image generation and editing model appeared anonymously on a public model-comparison leaderboard under that name before launch, people loved the results, and the nickname stuck. Officially the first version is Gemini 2.5 Flash Image, and the later Nano Banana Pro is Gemini 3 Pro Image. Google itself now uses the nickname in its own apps and developer tools.

All of these are image models. An image model takes a text prompt, optionally with reference images, and returns a still picture. A video model returns a sequence of frames with motion, and often audio. Google’s video model family is Veo, which is a different product from Nano Banana.

So when someone says “nano banana video”, they usually mean one of three things:

• A workflow: generate keyframes with Nano Banana, then animate them with an image-to-video model such as Veo or Kling. This is the honest, useful meaning and the focus of this article.
• A third-party website: several sites use “nano banana” in their domain and offer video generation. They are not Google products. The video itself comes from other models, sometimes several at once, with Nano Banana possibly used for the first frame.
• A misunderstanding: the belief that Google ships a Nano Banana model that outputs clips. It does not, at least as of October 2026.

None of this makes the combination less valuable. Nano Banana is one of the best tools available for producing the stills that a video model starts from, and the quality of that start frame decides much of the quality of the clip.

Can Nano Banana make video on its own?

No. Every Nano Banana version you can call today returns images only. What you can do is use those images as input to an image-to-video model: a video model that takes a still picture as its first frame (some also accept a last frame) plus a text prompt describing the motion, and generates the in-between frames.

Why bother with two models instead of typing a prompt straight into a text-to-video model? Because the split plays to each model’s strengths:

• Control over the look. A still is cheap to regenerate. You can iterate on composition, lighting, wardrobe and framing for a few credits per attempt before spending anything on motion.
• Character consistency. Nano Banana accepts reference images, so the same face, outfit or product appears identically across every shot. Text-to-video models tend to drift between clips, and a fixed start frame anchors identity.
• Text and branding. Nano Banana Pro renders legible text, which matters for product shots, signage and title cards.
• Edits. Nano Banana can edit an existing photo with a plain-language instruction (“make it dusk”, “remove the car”), so a real product photo can become the start of a clip.

Think of the image model as the cinematographer and art director, and the video model as the camera operator and actors. A nano banana video is good when both jobs are done well.

Which Nano Banana model should you use for keyframes?

There are several Gemini image models under the Nano Banana name, plus a few non-Google alternatives worth knowing for specific tasks. Here is how they compare for video keyframe work on Unicorn API:

Model	Reference images	Resolution	Aspect ratios	Credits per image
Nano Banana Pro	Up to 8	1K, 2K or 4K	16:9, 1:1, 9:16, 21:9, 4:3	18 (1K or 2K), 24 (4K)
Nano Banana	Up to 3	Set via image_size	16:9, 1:1, 9:16, 21:9, 4:3	4
Nano Banana 2 Lite	Up to 10	1K	16:9, 1:1, 9:16, 21:9, 4:3, 3:2, 2:3	4
GPT Image 1.5	Up to 4	Low, medium or high quality	1:1, 3:2, 2:3	3, 4 or 22
Flux 3 Klein Edit Image	Input images via image_urls	1K, 2K or 4K	Set via aspect_ratio	1.6, 2.5 or 6.5

Nano Banana Pro for hero frames

Nano Banana Pro, released in November 2025, is the default choice when a frame will be seen on screen for any length of time. It renders at up to 4K, renders text accurately and accepts up to 8 reference images, which is enough for a character sheet, a product shot and a location photo together. For video, 2K is usually the sweet spot: a 1080p frame is 1920 by 1080 pixels, so a 2K still already covers it, and 1K and 2K cost the same 18 credits. Reserve 4K for frames you plan to crop or slowly zoom into.

Nano Banana 2 Lite for drafts and storyboards

Nano Banana 2 Lite, released in July 2026, is the cheapest Gemini image model at 4 credits per image. It is limited to 1K, but it accepts up to 10 reference images, which is the most of any model in the table. Use it to rough out a whole storyboard, test camera angles and get sign-off before regenerating the final frames with Pro. It also adds 3:2 and 2:3 ratios. If you want the full picture of the Nano Banana 2 line, see our Nano Banana 2 AI image generator guide.

The original Nano Banana for quick edits

The original Nano Banana from August 2025 still earns its place for fast, natural-looking edits at 4 credits. It takes up to 3 reference images, which is fine when you are adjusting a single photo rather than building a cast.

Where the alternatives fit

GPT Image 1.5 is strong at following prompts literally and can output a transparent background, but it only supports 1:1, 3:2 and 2:3, so it is a poor fit for 16:9 or 9:16 video frames. Flux 3 Klein Edit Image is a very cheap way to touch up an existing keyframe (1.6 credits at 1K), and Flux 3 Klein Text to Image is a cheap way to generate background plates or textures.

How to make a nano banana video, step by step

This is the workflow that holds up in practice, whether you click through a browser or script it.

1. Write a shot list. One line per shot: subject, action, setting, camera angle, and how long it should last. Most image-to-video models produce short clips, so plan a sequence of short shots rather than one long take.
2. Lock your subject with a reference image. Generate a clean, well-lit portrait of your character or product on a plain background. This becomes the reference you pass to every later request. Our guide to consistent AI characters covers this step in more depth.
3. Pick the aspect ratio of the final video now. Use 16:9 for YouTube and landscape ads, 9:16 for Reels, Shorts and TikTok. Generate every keyframe at that ratio. Cropping later throws away resolution and often cuts off the subject.
4. Generate one keyframe per shot. Pass the reference image in ref_images and describe the shot. Keep the frame readable: one clear subject, a defined background, a pose that suggests the motion to come.
5. Fix details in the still. A wrong hand, stray text or a cluttered background should be fixed with an edit before animation. Retrying a still costs a few credits; retrying a clip costs much more.
6. Animate each keyframe. Send the still to an image-to-video model with a motion prompt that describes camera movement and action (“slow push-in, she lifts the cup and smiles”), not appearance. The image already defines appearance; repeating it in the prompt can make the model redraw things.
7. Edit and add sound. Trim the clips, put them in order and add music, voiceover or sound effects in your editor. If your video model does not produce audio, separate audio models can.

If you want to try the keyframe step without writing code first, the image playground lets you run each Nano Banana model in the browser with the same options as the API.

Generating nano banana video keyframes through the API

Once a workflow works by hand, scripting it saves hours, especially when a client asks for “the same thing, but in nine variations”. Media models on Unicorn API all follow the same pattern: create a job with POST /v1/jobs, poll GET /v1/jobs/{id} until status is "succeeded", then read the URL from outputs[0].url. You will need an API key from the keys page.

A single keyframe with curl

curl -X POST https://api.unicornapi.net/v1/jobs \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-pro",
    "input": {
      "prompt": "A small round white robot with blue eyes, standing on a wooden desk, morning light through a window, cinematic, shallow depth of field",
      "resolution": "2K",
      "aspect_ratio": "16:9"
    }
  }'

# Then poll with the id from the response:
curl https://api.unicornapi.net/v1/jobs/JOB_ID \
  -H "Authorization: Bearer $UNICORN_API_KEY"

A full storyboard in Python

This script generates a reference portrait, then four shots that all reuse it, so the robot looks the same in every frame. It prints the image URLs, ready to hand to an image-to-video model.

import os
import time
import requests

API = "https://api.unicornapi.net"
HEADERS = {
    "Authorization": f"Bearer {os.environ['UNICORN_API_KEY']}",
    "Content-Type": "application/json",
}

def run_job(model, inp, poll_seconds=3):
    """Create a job, wait for it to finish, return the first output URL."""
    resp = requests.post(f"{API}/v1/jobs", headers=HEADERS,
                         json={"model": model, "input": inp})
    resp.raise_for_status()
    job = resp.json()

    while job["status"] not in ("succeeded", "failed"):
        time.sleep(poll_seconds)
        poll = requests.get(f"{API}/v1/jobs/{job['id']}", headers=HEADERS)
        poll.raise_for_status()
        job = poll.json()

    if job["status"] != "succeeded":
        raise RuntimeError(f"Job {job['id']} ended with status {job['status']}")
    return job["outputs"][0]["url"]

CHARACTER = ("a small round white robot with glowing blue eyes "
             "and a thin antenna, matte plastic finish")

# 1. Reference image: plain background, full body, even light
ref_url = run_job("nano-banana-pro", {
    "prompt": f"Character reference sheet of {CHARACTER}, front view, "
              "full body, plain light grey background, soft even studio light",
    "resolution": "2K",
    "aspect_ratio": "16:9",
})
print("reference:", ref_url)

# 2. One keyframe per shot, all using the same reference image
shots = [
    "lying on a wooden desk with its eyes dim, early morning, wide shot",
    "standing up on the desk, eyes glowing, medium shot, warm window light",
    "walking toward a rain-streaked window, seen from behind, medium shot",
    "close-up, pressing one hand against the window glass, city lights outside",
]

for i, shot in enumerate(shots, start=1):
    url = run_job("nano-banana-pro", {
        "prompt": f"The same robot from the reference image, {CHARACTER}, {shot}. "
                  "Cinematic, consistent lighting, 35mm film look",
        "ref_images": [ref_url],
        "resolution": "2K",
        "aspect_ratio": "16:9",
    })
    print(f"shot {i}:", url)

A few notes on the script. It repeats the character description in every prompt as well as passing the reference image; doing both gives noticeably steadier results than either alone. Swap "nano-banana-pro" for "nano-banana-2-lite" and drop the resolution field to draft the whole storyboard at 4 credits per frame. Output URLs are best downloaded and stored on your side rather than treated as permanent.

Handing the frames to a video model

The animation step uses exactly the same job pattern: a different model id, an image URL as input and a motion prompt. Input field names vary between video models, so open the model’s page in the model catalog to see what it expects, and check the API documentation for details. For a comparison of the video models themselves, read our AI video generator tool guide. If you work inside Cursor or Claude Code, the MCP setup lets an agent run this whole pipeline for you.

How much does a nano banana video cost?

Cost has two parts: the keyframes and the animation. The keyframe side is easy to predict, because every image model on Unicorn API has a fixed credit price per output.

Credits cost $0.01 each on a monthly plan (Starter $25, Creator $50, Studio $100, Pro $200 per month). Top-up credits cost $0.015 each and never expire. That puts keyframes at:

• Nano Banana or Nano Banana 2 Lite: 4 credits, so $0.04 on a plan or $0.06 from top-ups.
• Nano Banana Pro at 1K or 2K: 18 credits, so $0.18 on a plan or $0.27 from top-ups.
• Nano Banana Pro at 4K: 24 credits, so $0.24 on a plan or $0.36 from top-ups.

The n field multiplies the price: asking for four variations of one shot at 2K with Pro costs 72 credits.

A worked example: a six-shot product clip

Say you are making a short vertical ad with six shots of a coffee mug. A realistic image budget looks like this:

Task	Model and setting	Count	Credits
Product reference shots	Nano Banana Pro, 2K	2	36
Storyboard drafts	Nano Banana 2 Lite	12	48
Final keyframes (two tries each)	Nano Banana Pro, 2K	12	216
Touch-up edits	Flux 3 Klein Edit Image, 2K	3	9 (2.5 rounds up to 3 each)
Total			309

That is $3.09 on a monthly plan, or about $4.64 from top-ups, for every still in the project. The Starter plan’s 2,500 credits would cover the image side of this project several times over.

The animation side depends entirely on which image-to-video model you pick, its clip length and resolution. Rather than guessing, ask the API: POST /v1/models/price returns the exact price of any request before you run it, so a script can total up a whole project before spending a credit. For more detail on the image side, our Nano Banana pricing breakdown covers every option. Plans and top-ups are listed on the pricing page.

Is any of this free?

Google offers some free Nano Banana image generation in its own Gemini apps and developer tools, subject to limits. That is fine for experimenting with keyframes by hand. Sites advertising a free nano banana video generator are a different matter: read what “free” covers, because limits on resolution, number of clips, watermarks or queue priority are common. And remember that the motion still comes from another model.

Common mistakes and trade-offs

Most disappointing results come from a handful of avoidable errors.

• Expecting the image model to plan the motion. A keyframe is a single moment. If you want a character to walk across the room, generate the frame where they start walking, with their weight shifted forward, not a static portrait.
• Mismatched aspect ratios. A 1:1 still sent to a 9:16 video model gets cropped or padded. Generate at the final ratio from the start.
• Busy frames. Crowds, intricate patterns and lots of small objects give video models more to get wrong, and they tend to warp or flicker. Simple, well-separated compositions animate cleanly.
• Re-describing the image in the motion prompt. Describe camera movement, action and pacing. Restating colours and clothing invites the video model to redraw them.
• Fixing problems in the clip instead of the still. If the hand looks wrong in the first frame, it will look wrong in every frame. Fix it in the image, where a retry costs 4 to 18 credits.
• Relying on text staying sharp. Nano Banana Pro renders text accurately in the still, but motion can still distort it. Keep on-screen text on frames with little movement, or add it in your editor.
• Not saving the reference image. Lose it and you lose your character. Store it with the project and reuse the same URL or file.

Subscription site or API?

All-in-one nano banana video sites are convenient: one prompt box, one subscription, a finished clip. The trade-off is control. You often cannot see which model rendered your clip, swap it for a better one, or pay only for what you use. Working through an API means a little more setup, but you choose the image model and the video model separately, see the exact price of each request, and can automate the whole pipeline. If you make more than the occasional clip, that control usually pays for itself.

The short version: Nano Banana does not make video, but it makes the best possible starting point for one. Get the stills right with Nano Banana Pro or the cheaper Nano Banana 2 Lite, hand them to the video model of your choice, and you have a nano banana video that looks deliberate rather than random.