Astra Synthetic Data Block

An Edge Impulse custom synthetic data block that generates images with GPT-6 Astra and GPT Image 2.5 and uploads them to your project. In Studio it appears as Astra Synthetic Data.

Custom synthetic data blocks require the Edge Impulse Enterprise plan. They run without input files: Studio passes parameters and environment variables to the container, and the container uploads what it generates through the Ingestion API.

transform.py calls the OpenAI Responses API with gpt-6-astra and the image_generation tool, using gpt-image-2.5-sunburst or gpt-image-2.5-flare. It checks that each result is a valid PNG, then uploads it to Edge Impulse. synthetic_blocks_tutorial.md walks through the code and how to extend it.

Requirements

• Edge Impulse CLI
• Docker Desktop
• An OpenAI API key with access to GPT-6 Astra and GPT Image
• An Edge Impulse project API key, from Dashboard > Keys

Check the tools are installed:

edge-impulse-blocks --version
docker version

Parameters

Studio field	Passed as	Default
OpenAI API Key	OPENAI_API_KEY environment variable	none
Prompt	--prompt	A photo of a factory worker wearing a hard hat
Label	--label	hard_hat
Number of images	--images	3
Image model	--image-model	gpt-image-2.5-sunburst
Upload to category	--upload-category (split, training, testing)	split

Studio also passes --synthetic-data-job-id. The block sends it as the x-synthetic-data-job-id header on every upload, which is what makes samples preview on the Synthetic data tab.

Each sample is uploaded with generated_by, image_model, and prompt metadata, plus revised_prompt when OpenAI returns one.

Test Locally

Synthetic data blocks don't work with edge-impulse-blocks runner, so build and run the container directly. Set OPENAI_API_KEY and EI_PROJECT_API_KEY in your shell first. -e NAME with no value passes each key through without writing it into the command.

docker build -t astra-synthetic-data .

Generate one image without uploading it. It is saved to output/:

docker run --rm \
  -e OPENAI_API_KEY \
  -v "$PWD/output:/app/output" \
  astra-synthetic-data \
  --prompt 'A photo of a factory worker wearing a hard hat' \
  --label hard_hat \
  --images 1 \
  --skip-upload

Generate one image and upload it to your project:

docker run --rm \
  -e OPENAI_API_KEY \
  -e EI_PROJECT_API_KEY \
  astra-synthetic-data \
  --synthetic-data-job-id 123456789 \
  --prompt 'A photo of a factory worker wearing a hard hat' \
  --label hard_hat \
  --images 1 \
  --image-model gpt-image-2.5-sunburst \
  --upload-category training

The job ID here is a placeholder, so check the sample in Data acquisition rather than the Synthetic data tab.

Publish

.ei-block-config is gitignored. On a fresh clone, run edge-impulse-blocks init first and choose the synthetic data block type. Then push:

edge-impulse-blocks push

In an Enterprise project, open Data acquisition > Synthetic data, pick Astra Synthetic Data, and generate a small batch to check the previews and labels.

Agent Skill

.github/skills/create-synthetic-data-block/SKILL.md is a GitHub Copilot skill for extending this block. Copilot loads skills from .github/skills when this folder is open. Run /create-synthetic-data-block in Copilot Chat, or name the skill in a prompt:

Use the create-synthetic-data-block skill to add a quality option to the Astra block.
Keep the GPT-6 Astra Responses API call and the synthetic data job ID header.

AGENTS.md holds the rules that apply to any change in this repository.

Further Reading

• Custom synthetic data blocks
• parameters.json reference
• Ingestion API
• GPT-6 Astra model
• OpenAI image generation
• Install and write Agent Skills