Images

LLM API provides dedicated OpenAI-compatible image endpoints for text-to-image generation and image editing. Generated images are returned as base64 data in data[].b64_json; the exact model, size, quality, and output format support is catalog-dependent.

Generate an image

POST /v1/images/generations accepts a JSON body:

curl -X POST "https://api.llmapi.ai/v1/images/generations" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A tiny bookstore in a rainy neon city",
    "size": "1024x1024",
    "quality": "low"
  }'

Set stream: true with partial_images when the model supports incremental renders. Those renders arrive as SSE events. Without streaming, wait for the complete JSON response and decode each data[].b64_json value locally.

{
  "created": 1715000000,
  "data": [{ "b64_json": "iVBORw0KGgoAAA..." }],
  "usage": {
    "input_tokens": 12,
    "output_tokens": 500,
    "total_tokens": 512
  }
}

For example, decode an image in Python:

import base64
from pathlib import Path

image = response.json()["data"][0]["b64_json"]
Path("generated.png").write_bytes(base64.b64decode(image))

Generation fields

Field	Description
`model`	Image model ID. The model catalog is authoritative for supported parameters.
`prompt`	Text description or edit instruction. Required for both generation and editing.
`size`	Requested dimensions, such as `1024x1024` or `auto`, where supported.
`quality`	`low`, `medium`, `high`, `xhigh`, `max`, or `auto`, depending on the model.
`n`	Number of images to generate, if supported.
`output_format`	`png`, `jpeg`, or `webp`, where supported.
`background`	`auto`, `opaque`, or `transparent`, where supported.
`output_compression`	JPEG/WebP compression value from 0 to 100.
`stream` / `partial_images`	Enable SSE and choose how many intermediate renders to emit.

Unsupported parameters are rejected by some models and ignored by others, so check the model capability data before relying on a parameter.

Edit an image

POST /v1/images/edits uses multipart/form-data. Upload one or more images under image[], and optionally upload a PNG mask:

curl -X POST "https://api.llmapi.ai/v1/images/edits" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=Turn the sky into a pink sunset" \
  -F "image[]=@photo.png"

Repeat image[] for multi-image editing. A mask identifies the editable area when the selected model supports masked edits; the mask should be a PNG with the dimensions expected by the model. The edit response uses the same data[].b64_json and usage shape as image generation.

curl -X POST "https://api.llmapi.ai/v1/images/edits" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -F "model=gpt-image-2" \
  -F "prompt=Replace the sign with a blank wooden sign" \
  -F "image[]=@street.png" \
  -F "mask=@sign-mask.png" \
  -F "output_format=png"

Streaming partial renders

For models that support image streaming, request partial renders explicitly:

{
  "model": "gpt-image-2",
  "prompt": "A watercolor lighthouse",
  "size": "1024x1024",
  "stream": true,
  "partial_images": 2
}

The server returns SSE events containing intermediate base64 image data. They are previews, not additional final images; use the final event/result for the completed asset. If a model does not advertise streaming support, use the normal synchronous request.

Notes

Image models are billed according to the selected model, quality, size, number of images, and output token usage. Start with a low-quality, small render while integrating.
Dedicated image endpoints are separate from chat-based image models. For models that return images through a multimodal chat response, use Chat Completions and see Image Generation.
The Models page shows current image models and capabilities.
Image data is returned inline as base64 rather than as a permanent gateway-hosted URL. Decode it promptly and store it in your own object storage if it must be retained.
Common failures include unsupported size or quality, oversized uploads, insufficient credits, rate limiting, and upstream provider errors.