# Gemini 3

Source: https://docs.prodia.com/models/gemini-3/

Gemini 3 is Google's higher-tier image family — the same multimodal lineage as [Nano Banana](https://docs.prodia.com/models/nano-banana/), but with selectable resolution up to 4K and stronger fidelity on complex prompts. Two variants are exposed through Prodia: *Gemini 3 Pro* for highest-quality output and *Gemini 3.1 Flash* for cost-efficient generation with optional Google Search grounding. Both accept text prompts, support image-to-image editing, and complete in 10–35 seconds depending on resolution.

> note:
>
> In the Prodia explorer these variants are surfaced as *Nano Banana Pro* and *Nano Banana 2*. The underlying job-type identifiers (`inference.gemini-3-pro.*` and `inference.gemini-3-1-flash.*`) follow Google's naming. The IDs in this page are what you send to the API.

### Architecture

Gemini 3 is a multimodal model: text and images share a single context window, and image generation runs through the same reasoning path as text generation. In practice this is what makes editing behave like instruction-following rather than diffusion noise blending — the model reads the input image as visual tokens, reasons about which regions the prompt describes, and rewrites only those pixels.

The Pro variant is the higher-capability tier, biased toward complex compositions, accurate text rendering, and detail retention at high resolutions. Flash is the smaller, faster sibling — same model family, lower cost per image, and uniquely able to ground generation in real-time Google Search results when `google_search` is enabled.

### Choosing a variant

| Variant | Best for                                                     | Generation time | Resolutions | Price                              |
| ------- | ------------------------------------------------------------ | --------------- | ----------- | ---------------------------------- |
| *Pro*   | High-fidelity output, hardest prompts, 4K detail             | \~10–12s        | 1K, 2K, 4K  | $0.15 (1K/2K), $0.30 (4K)          |
| *Flash* | Cost-efficient generation, search grounding, batch workloads | \~30–35s        | 1K, 2K, 4K  | $0.08 (1K), $0.12 (2K), $0.16 (4K) |

*Gemini 3 Pro* — flagship image model. Accurate text rendering, photorealistic detail, and consistent multi-object composition. Img2img accepts up to 3 reference images.

*Gemini 3.1 Flash* — smaller, cheaper, and supports `google_search` grounding for prompts that require real-world facts (logos, product designs, current events). Img2img accepts up to 14 reference images, useful for multi-image composition or character-consistency workflows across larger image sets.

### When to use Gemini 3

- *4K-resolution output* — neither [Nano Banana](https://docs.prodia.com/models/nano-banana/) nor [FLUX.2](https://docs.prodia.com/models/flux-2/) generates above 2K natively. Use Gemini 3 when you need print-quality detail
- *Accurate text rendering at high resolution* — both variants render text reliably; for native vector text output, see [Recraft V4](https://docs.prodia.com/models/recraft-v4/)
- *Multi-image composition with up to 14 inputs* (Flash) — combining a subject across many reference frames, or fusing several scenes into one
- *Search-grounded generation* (Flash only) — set `google_search: true` to pull in real-world references during generation. Useful for logos, products, locations, or anything that benefits from current factual context
- *Predictable resolution-tier pricing* — the only billing parameter is `resolution`. No per-step or per-pixel cost surprises

For natural-language editing at flat per-job pricing, [Nano Banana](https://docs.prodia.com/models/nano-banana/) is a strong alternative. For instruction-guided edits with arbitrary aspect ratios, see [FLUX.1 Kontext](https://docs.prodia.com/models/flux-kontext/).

### Job types

| Job type                                | Description                                                      | ETA   |
| --------------------------------------- | ---------------------------------------------------------------- | ----- |
| `inference.gemini-3-pro.txt2img.v1`     | Gemini 3 Pro text-to-image                                       | \~10s |
| `inference.gemini-3-pro.img2img.v1`     | Gemini 3 Pro image-to-image (up to 3 inputs)                     | \~12s |
| `inference.gemini-3-1-flash.txt2img.v1` | Gemini 3.1 Flash text-to-image, optional Google Search grounding | \~30s |
| `inference.gemini-3-1-flash.img2img.v1` | Gemini 3.1 Flash image-to-image (up to 14 inputs)                | \~35s |

> note:
>
> 4K Pro and 2K/4K Flash jobs can take 30s+ to complete. For long-running generations, use the [async polling API](https://docs.prodia.com/guides/polling-async-jobs/) instead of holding the connection open.

### Parameters

*Common to all four job types:*

- `prompt` (required) — text description of the desired output, 1–5,000 characters. For img2img this is the editing instruction
- `aspect_ratio` — one of `1:1` (default), `2:3`, `3:2`, `3:4`, `4:3`, `4:5`, `5:4`, `9:16`, `16:9`, `21:9`
- `resolution` — `1K` (default), `2K`, or `4K`. Use uppercase K. Pricing tier is determined by this field
- `include_messages` — boolean, default `false`. When `true`, the multipart response also returns `message.txt` parts containing the model's natural-language reasoning

*`gemini-3-pro.img2img.v1`:*

- `images` — optional array of 1–3 input image filenames sent as multipart `input` parts. When omitted, a single attached input is used implicitly

*`gemini-3-1-flash.txt2img.v1` and `img2img.v1`:*

- `google_search` — boolean, default `false`. Enables Google Search grounding so the model can incorporate real-world references (logos, products, locations) into the output
- `images` — `img2img` only, optional array of 1–14 input image filenames

### Prompting tips

- *Describe the change, not the whole scene.* For img2img write the diff — "add a small wooden tag with a leather string" — rather than re-describing the input. The model already sees it
- *Anchor preservation explicitly.* "Keep everything else exactly the same" reduces drift in untouched regions, especially for subtle edits at 1K
- *Use `4K` only when you need it.* 4K Pro is 2x the price of 1K/2K, and Flash 4K takes 30s+. For thumbnails and web use, 1K is usually enough
- *Enable `google_search` for fact-bound prompts* — brand logos, product replicas, real-world architecture. Skip it for purely creative prompts where grounding adds latency without value
- *Pass multiple images deliberately.* When using multi-image inputs, refer to them by position in the prompt — "the subject from the first image, in the setting from the second" — and make sure the multipart filenames match the order in `images`
- *Pick `aspect_ratio` up front.* The default is `1:1`. For social, web, or phone, set `9:16` or `16:9` rather than upscaling or cropping after the fact

### Examples

Gemini 3 Pro text-to-image at 16:9, 1K resolution:

```json
{
  "type": "inference.gemini-3-pro.txt2img.v1",
  "config": {
    "prompt": "a cute red panda eating bamboo on a mossy log, soft natural light, photorealistic",
    "aspect_ratio": "16:9",
    "resolution": "1K"
  }
}
```

![Gemini 3 Pro txt2img — red panda eating bamboo on a mossy log](https://docs.prodia.com/llms/assets/fce243c00562-gemini-3-pro-txt2img.jpg)

Same image edited with `img2img.v1` — adds a wooden name tag while preserving the pose, fur, lighting, and background:

```json
{
  "type": "inference.gemini-3-pro.img2img.v1",
  "config": {
    "prompt": "Add a small wooden tag with a leather string around the red panda's neck. Keep everything else exactly the same.",
    "aspect_ratio": "16:9",
    "resolution": "1K"
  }
}
```

![Gemini 3 Pro img2img — same panda with a wooden name tag added](https://docs.prodia.com/llms/assets/dfcfec6c768f-gemini-3-pro-img2img.jpg)

Gemini 3.1 Flash text-to-image at 2K, generating a multi-panel infographic with rendered text:

```json
{
  "type": "inference.gemini-3-1-flash.txt2img.v1",
  "config": {
    "prompt": "a hyper-realistic infographic poster about coffee brewing methods, clean modern design, soft pastel background",
    "aspect_ratio": "3:4",
    "resolution": "2K"
  }
}
```

![Gemini 3.1 Flash txt2img — coffee brewing methods infographic](https://docs.prodia.com/llms/assets/c1779e2672de-gemini-3-1-flash-txt2img.jpg)

Gemini 3.1 Flash img2img — same red panda transformed into a winter scene:

```json
{
  "type": "inference.gemini-3-1-flash.img2img.v1",
  "config": {
    "prompt": "Make this look like a winter scene with snow falling and snow covering the moss and ferns. Keep the red panda in the same pose.",
    "aspect_ratio": "16:9",
    "resolution": "1K"
  }
}
```

![Gemini 3.1 Flash img2img — same panda in a snowy winter scene](https://docs.prodia.com/llms/assets/03745cf8e4d3-gemini-3-1-flash-img2img.jpg)

Search-grounded generation with Flash — pull a real-world reference into the output:

```json
{
  "type": "inference.gemini-3-1-flash.txt2img.v1",
  "config": {
    "prompt": "a product render of the original 1984 Macintosh 128K on a clean studio backdrop, accurate proportions, soft top lighting",
    "aspect_ratio": "1:1",
    "resolution": "2K",
    "google_search": true
  }
}
```

### Guides

[Generating Images](https://docs.prodia.com/guides/generating-images/) — Step-by-step guide for generating images with Prodia.

[Transforming Images](https://docs.prodia.com/guides/transforming-images/) — Use img2img to transform or edit an existing image.

[Polling Async Jobs](https://docs.prodia.com/guides/polling-async-jobs/) — Use the async API for long-running 4K generations.