dubery-fidelity-prompt

To generate JSON prompts with locked product identity, use a fidelity-first approach that loads product specs from a dedicated JSON file and treats the product as a fixed asset while applying variable scene parameters. This ensures consistent rendering across multiple product images without manual prompt crafting.

The best way to structure image prompts for Vertex AI is using a v3 JSON schema that includes sections for product_fidelity, scene_variables, render_quality, and api_parameters. This schema-driven method outputs a formatted text prompt and a companion config file containing image_input data for the pipeline.

Yes, you can use scene randomizers to create variations without losing fidelity. The system accepts scene parameters from randomizers or direct input, treating the scene as a variable asset while strictly enforcing product specs loaded from sidecar metadata to preserve the core product identity.

Before generating prompts, you need product specifications defined in a contents/assets/product-specs.json file and any per-prodref sidecar metadata. These files establish the locked product identity and physical interaction rules required for the rendering pipeline.

A companion config JSON file is needed to provide the image_input reference required by Vertex AI rendering pipelines. The prompt generator outputs both a formatted text prompt and this config file, ensuring the rendering engine can map the scene variables to the locked product reference image.