Image Generation
Input: Text · Image
Gemini Nano Banana
For detailed usage, refer to the Nano Banana API Guide.
| Provider | Model ID | Capabilities | Endpoint | Price (per 1M Tokens) | Launch Date | Deprecation Date | Capacity | Supported Regions | Notes |
|---|---|---|---|---|---|---|---|---|---|
| Vertex AI | turing/gemini-2.5-flash-image | Input: Output: Tools: Not supported | v1/chat/completions | Input: $0.3Output: $30 | 2025-09-12 | - | Vertex AI multi-project | China region Europe region North America region | - |
| Vertex AI | turing/gemini-3-pro-image | Input: Output: Tools: Not supported | v1/chat/completions | Input: $2Output: $60 | 2025-09-12 | - | Vertex AI multi-project | China region Europe region North America region | - |
| Vertex AI | turing/gemini-3.1-flash-lite-image | Input: Output: Tools: Not supported | v1/chat/completions | Input: $0.25Text output: $1.50Image output: $30 | 2026-07-01 | - | Vertex AI multi-project | China region Europe region North America region | Nano Banana 2 Lite, up to 1K output |
Azure OpenAI
| Provider | Model ID | Capabilities | Endpoint | Price (per image) | Launch Date | Deprecation Date | Capacity | Supported Regions | Notes |
|---|---|---|---|---|---|---|---|---|---|
| Azure | turing/gpt-image-2.5-sunburst | Input: Output: Tools: Not supported | v1/images/generationsv1/images/edits | Billed per token (per 1M tokens):Text input: $5.00 / cached $1.25Image input: $8.00 / cached $2.00Image output: $30.00 | 2026-09-11 | - | Global | China region Europe region North America region | GPT Image 2.5 base model, quality and editing precision first; quality adds xhigh / max. See GPT-image-2.5 |
| Azure | turing/gpt-image-2.5-flare | Input: Output: Tools: Not supported | v1/images/generationsv1/images/edits | Billed per token (per 1M tokens):Text input: $5.00 / cached $1.25Image input: $8.00 / cached $2.00Image output: $30.00 | 2026-09-11 | - | Global | China region Europe region North America region | GPT Image 2.5 small model, speed first, image quality comparable to GPT-image-2; same parameters as Sunburst |
| Azure | turing/gpt-image-2 | Input: Output: Tools: Not supported | v1/images/generations | Billed per token (per 1M tokens):Text input: $5.00 / cached $1.25Image input: $8.00 / cached $2.00Image output: $30.00 | 2026-04-23 | - | Global | China region Europe region North America region | Public preview. Supports arbitrary resolution (4K) and optional transparent background |
| Azure | turing/gpt-image-1 | Input: Output: Tools: Not supported | v1/images/generations | Low Quality: $0.011 (1024x1024) $0.016 (1024x1536) $0.016 (1536x1024)Medium Quality: $0.042 (1024x1024) $0.063 (1024x1536) $0.063 (1536x1024)High Quality: $0.167 (1024x1024) $0.25 (1024x1536) $0.25 (1536x1024) | 2025-06-19 | - | Global | China region Europe region North America region | - |
| Azure | turing/dall-e-3 | Input: Output: Tools: Not supported | v1/images/generations | $0.04 (1024x1024)$0.08 (1024x1536)$0.08 (1536x1024) | 2025-04-22 | 2026-02-18 | Global | China region Europe region North America region | - |
GPT-image-2.5 Technical Specifications
Both 2.5 models share GPT-image-2's endpoints and size rules. The differences: quality adds the xhigh / max tiers (default auto), resolutions above 2560x1440 are labelled experimental by OpenAI; OpenAI's docs list background="transparent" as supported, but on Azure it does not take effect yet (measured). See Image Generation API → GPT-image-2.5.
GPT-image-2 Technical Specifications
Public preview. See Image Generation API → GPT-image-2 for details.
Resolution Support
gpt-image-2 is no longer limited to a fixed set of resolutions. You can pass any size string in the format <w>x<h>, subject to the following constraints:
- Both sides must be integer multiples of 16 pixels
- Longest side ≤ 3840 px (4K supported)
- Aspect ratio ≤ 3:1
- Total pixel count between
655,360and8,294,400
You may also pass size="auto" to let the model choose the size automatically.
Image Generation Capabilities
| Capability | Description |
|---|---|
| Text to image | POST /v1/images/generations |
| Transparent background | background="transparent" (requires output_format="png") |
ByteDance Volcano
| Provider | Model ID | Capabilities | Endpoint | Price (per image) | Launch Date | Deprecation Date | Capacity | Supported Regions | Notes |
|---|---|---|---|---|---|---|---|---|---|
| ByteDance Volcano | doubao-seedream-4-0-250828 | Input: Output: Tools: Not supported | v1/images/generations | ¥0.2 per image | 2025-12-12 | - | ByteDance Volcano | China region | - |
| ByteDance Volcano | doubao-seedream-4-5-251128 | Input: Output: Tools: Not supported | v1/images/generations | ¥0.25 per image | 2025-12-12 | - | ByteDance Volcano | China region | - |
| ByteDance Volcano | doubao-seedream-5-0-260128 | Input: Output: Tools: Web search | v1/images/generations | ¥0.22 per image | 2026-01-28 | - | ByteDance Volcano | China region | This is Seedream 5.0-lite. Supports web search (web_search), streaming and batch generation (n). The older id doubao-seedream-5-0-lite-260128 still works |
| ByteDance Volcano | doubao-seedream-5-0-pro-260628 | Input: Output: Tools: Web search | v1/images/generations | single image: ≤2.61M px ¥0.30 / >2.61M px ¥0.60layer split: ≤2.61M px ¥0.15 / >2.61M px ¥0.30input ¥0.02 per image (first one free) | 2026-06-28 | - | ByteDance Volcano | China region | Flagship tier, supports interactive editing, layer split and multi-image fusion. Supports neither streaming nor batch generation (n). Billed per image, tiered by generation scene and by that image's actual output size |
Seedream Technical Specifications
Resolution Support
| Model | Supported Resolutions |
|---|---|
| doubao-seedream-4-0-250828 | 1K, 2K, 4K |
| doubao-seedream-4-5-251128 | 2K, 4K |
| doubao-seedream-5-0-260128 | 2K, 3K, 4K |
| doubao-seedream-5-0-pro-260628 | 1K, 1.5K, 2K (layer split also accepts auto) |
- Aspect ratio range: [1/16, 16]
- Supported image formats: JPEG, PNG, WEBP, BMP, TIFF, GIF (
doubao-seedream-5-0-260128additionally supports HEIC and HEIF)
Image Generation Capabilities
- Text to image: Generate high-quality images from text descriptions
- Image to image: Generate new images based on a reference image
- Multi-image fusion: Supports 2–14 reference images combined with a text prompt to generate images
- Batch image generation: Up to 15 images (number of input images + number of generated images ≤ 15)
The n Parameter and Batch Generation
When using a Seedream model, the n parameter is automatically converted to a batch image generation parameter. doubao-seedream-5-0-pro-260628 is the exception: it does not support batch generation, so n is ignored and a single image is always returned:
# When n=3 is set, the platform automatically converts it to:
payload["sequential_image_generation"] = "auto"
payload["sequential_image_generation_options"] = {"max_images": 3}
sequential_image_generation accepted values:
"auto": The model automatically decides whether to return multiple images and how many"disabled": Generate only a single image
Parameter Differences Across the Seedream 5.0 Versions
| Parameter | doubao-seedream-5-0-260128 | doubao-seedream-5-0-pro-260628 |
|---|---|---|
image (image-to-image) | Supported | Supported |
watermark | Supported | Supported |
optimize_prompt_options | Supported | Supported |
output_format (png/jpeg) | Supported | Supported |
background (opaque/transparent) | Not supported | Supported (image-to-image only, and the input image must carry an alpha channel) |
layer_decomposition | Not supported | Supported (layer split, single input image only) |
tools (web search) | Supported | Not supported |
guidance_scale | Not supported | Not supported |
stream | Supported | Not supported |
sequential_image_generation / n | Supported | Not supported |
Parameters the upstream does not support are not forwarded, so passing them has no effect
(stream: true is rejected during request validation instead).
For web search, pass tools: [{"type": "web_search"}]; the response reports the number of
lookups under usage.tool_usage.
Layer Split (doubao-seedream-5-0-pro-260628 only)
Pass layer_decomposition: true to split a single input image into one base image plus up to 16
transparent PNG layers.
imagebecomes required and accepts exactly one image (png / jpeg, total pixels [262144, 36M])promptbecomes optional: omit it and the model detects and splits every main element on its own; supply it to target specific elements (<bbox>normalized coordinates are supported)- In this scene
sizeaccepts the tiers1K/1.5K/2K/autoonly. When omitted the platform sendsauto, because the1024x1024schema default is a pixel value the upstream rejects output_formatcontrols the base image only; layers are alwayspng- The response
dataarray holds both the base image and the layers, told apart byz_index(0 for the base image), and additionally carriesbounding_box.absolute/bounding_box.normalized/name/description - A layer can be fed back in as an input image for further editing; to keep the transparent
background, set
background: "transparent"together withoutput_format: "png"
Billing counts the images in usage.generated_images, the base image included, and tiers each one by
its own output size — the layer-split rate is half the single-image rate. A measured auto-split of 14
outputs (a 1760×2176 base plus 13 layers) is billed ¥2.25.
Alibaba Model Studio
| Vendor | Model ID | Capabilities | endpoint | Price (per image) | Launch date | Planned retirement | Capacity | Available regions | Notes |
|---|---|---|---|---|---|---|---|---|---|
| Alibaba Model Studio | wan2.7-image | Input: Output: Tools: Not supported | v1/images/generations | ¥0.20 per image | 2026-04-01 | - | Alibaba Model Studio | China region | Wan 2.7, up to 2K |
| Alibaba Model Studio | wan2.7-image-pro | Input: Output: Tools: Not supported | v1/images/generations | ¥0.50 per image | 2026-04-01 | - | Alibaba Model Studio | China region | Wan 2.7 flagship tier, 4K for text-to-image |
Wan Technical Specifications
Resolution Support
| Model | Supported resolutions |
|---|---|
| wan2.7-image | 1K, 2K, or WIDTHxHEIGHT |
| wan2.7-image-pro | 1K, 2K, 4K (4K for text-to-image only), or WIDTHxHEIGHT |
When size is given as WIDTHxHEIGHT, the platform rewrites it into the WIDTH*HEIGHT form the upstream requires, so callers can keep writing 1024x1024 as they would for OpenAI.
Image Generation Capabilities
- Text to image: Generate an image from a text description
- Image to image: Pass one or more reference images (URL or base64) through
image - Sequential (group) generation:
enable_sequential: true
Supported Parameters
Beyond prompt, model, image and size, the following parameters are forwarded upstream as-is:
| Parameter | Type | Description |
|---|---|---|
n | integer | With sequential mode off, the number of images to generate, 1–4 (default 1); with it on, the maximum number of images, 1–12 (default 12) |
enable_sequential | boolean | Enable group-image output mode. Default false |
thinking_mode | boolean | Enable thinking mode. Default true; only effective with sequential mode off and no image input, and it increases latency |
color_palette | array | Custom color theme: 3–10 {hex, ratio} entries whose ratio values must total 100.00%. Only available with sequential mode off |
bbox_list | array | Bounding boxes for interactive editing, given in input-image order |
watermark | boolean | Add an "AI generated" watermark. Default false |
seed | integer | Random seed in [0, 2147483647] |
Billing counts only the images actually returned (failed calls are not charged). With sequential mode on, omitting n lets the upstream generate up to 12 images — always set n explicitly. See Alibaba Model Studio billing for prices.