Image Generator Service | BetterAI Docs

Source: src/image_generation/image_generator_service.py. This page's original notebook was empty; the content was written from the code.

#Overview

The ImageGeneratorService class covers four steps, in this order: building the parts (text and images), configuring the generation, calling the model and parsing the responses (images, text and usage metrics).

#Constructor

Parameter	Type	Description
`client`	`genai.Client` (optional)	Gemini client already initialized. If `None`, a new one is created with `GeminiClient().get_client()`.
`content_config`	`dict` (optional)	Overrides values of `DEFAULT_CONTENT_CONFIG` (model, `temperature`, `top_p`, `aspect_ratio`, `number_of_images`, etc.).

The final configuration is the merge of DEFAULT_CONTENT_CONFIG with content_config: missing keys inherit the default value. Any failure during initialization raises RuntimeError.

#Methods

#build_parts

Builds the list of types.Part sent to Gemini. The text is made of BASE_PROMPT, followed by [TASK] with the user_prompt and, if instructions is set, by [USER_CONTEXT]. Each reference image in images is appended as a bytes part.

An empty user_prompt, or one that is not a str, raises an error (ValueError, re-raised as RuntimeError). The method also stores the prompt, the instructions and the image count in self.text_input, for later use in parse_responses.

#generate_config

Returns a types.GenerateContentConfig with temperature, top_p and max_output_tokens from the configuration, response_modalities equal to ["IMAGE", "TEXT"] and ImageConfig with the aspect_ratio.

#call_model

Calls client.models.generate_content once for each requested image, with the same prompt, and returns the list of responses. The number of calls is content_config["number_of_images"].

Cost and limits. The current Gemini SDK does not configure the number of images per call; this is why generating several images is done with several calls. Avoid high values of number_of_images to avoid exceeding rate limits or raising the cost.

#parse_responses

Walks through the responses and aggregates the result into a single dictionary:

Key	Content
`text_input`	Prompt, instructions and image count (only present after `build_parts`).
`text_responses`	List with the texts returned by the model.
`images`	List of dictionaries with `mime_type` and `data` (bytes).
`generate_config`	The final configuration used for the generation.
`usage_metadata`	One entry per response, with `prompt_tokens`, `output_tokens` and `total_tokens`, or `None` when the response carries no metrics.

#Usage example

from src.image_generation.image_generator_service import ImageGeneratorService

# Create the service with a partial configuration
service = ImageGeneratorService(content_config={"number_of_images": 2})

# 1. Build the parts, 2. configure, 3. call the model, 4. parse
parts = service.build_parts(user_prompt="Um grifo renascentista")
config = service.generate_config()
responses = service.call_model(parts, config)
result = service.parse_responses(responses)

print(len(result["images"]))

#Notes

Input image type. In build_parts, all reference images are sent with mime_type image/jpeg, even when the original file is a PNG.

See also Config, Gemini Client and Module.