NVIDIA NIM Generate Image And Caption

arazzo: 1.0.1
info:
  title: NVIDIA NIM Generate Image And Caption
  summary: Generate an image from a text prompt, then caption the generated image with a vision-language model.
  description: >-
    A generate-then-verify visual flow. A visual generative NIM (FLUX.1,
    SDXL, Edify) produces an image from a text prompt and returns it as a
    base64-encoded artifact. That artifact is then fed straight back into a
    vision-language NIM as a base64 data-URI image_url so the VLM can describe
    what was actually rendered, giving an automatic caption and a check that the
    output matched the prompt. Every step spells out its request inline so the
    flow can be read and executed without opening the underlying OpenAPI
    description.
  version: 1.0.0
sourceDescriptions:
- name: imageGenerationApi
  url: ../openapi/nvidia-nim-image-generation-api-openapi.yml
  type: openapi
- name: visionApi
  url: ../openapi/nvidia-nim-vision-api-openapi.yml
  type: openapi
workflows:
- workflowId: generate-image-and-caption
  summary: Generate an image, then describe it with a VLM to produce a caption.
  description: >-
    Generates an image from a prompt with a visual generative model, then passes
    the returned base64 artifact to a vision-language model for captioning.
  inputs:
    type: object
    required:
    - apiKey
    - publisher
    - model
    - prompt
    properties:
      apiKey:
        type: string
        description: NVIDIA developer API key (nvapi-...) sent as a Bearer token.
      publisher:
        type: string
        description: Image model publisher path segment (e.g. black-forest-labs).
      model:
        type: string
        description: Image model slug path segment (e.g. flux.1-schnell).
      prompt:
        type: string
        description: Text prompt describing the image to generate.
      visionModel:
        type: string
        description: Vision-language model id used to caption the generated image.
        default: meta/llama-3.2-90b-vision-instruct
  steps:
  - stepId: generateImage
    description: >-
      Generate an image from the text prompt with the selected publisher/model
      visual generative NIM.
    operationId: generateImage
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    - name: publisher
      in: path
      value: $inputs.publisher
    - name: model
      in: path
      value: $inputs.model
    requestBody:
      contentType: application/json
      payload:
        prompt: $inputs.prompt
        mode: text-to-image
        steps: 50
        cfg_scale: 5
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      imageBase64: $response.body#/artifacts/0/base64
      finishReason: $response.body#/artifacts/0/finishReason
      seed: $response.body#/artifacts/0/seed
  - stepId: captionImage
    description: >-
      Caption the freshly generated image by passing its base64 artifact as a
      data-URI image_url into a vision-language model.
    operationId: createVisionChatCompletion
    parameters:
    - name: Authorization
      in: header
      value: Bearer $inputs.apiKey
    requestBody:
      contentType: application/json
      payload:
        model: $inputs.visionModel
        messages:
        - role: user
          content:
          - type: text
            text: Write a one-sentence caption describing this generated image.
          - type: image_url
            image_url:
              url: $steps.generateImage.outputs.imageBase64
        max_tokens: 64
        temperature: 0.3
        stream: false
    successCriteria:
    - condition: $statusCode == 200
    outputs:
      caption: $response.body#/choices/0/message/content
      finishReason: $response.body#/choices/0/finish_reason
  outputs:
    imageBase64: $steps.generateImage.outputs.imageBase64
    caption: $steps.captionImage.outputs.caption