Skip to content
GenMediaOpen workspace
Browse documentationCandidates

Generate on fal or import finished outputs

Candidate creation has two paths. GenMedia can expand prompts into requests for regular, private, or ephemeral fal endpoints, or it can import finished media when the outputs already exist or the Evaluation Task compares treatment arms rather than endpoints.

Generate from fal endpoints

  1. 1

    Choose a project, modality, comparison design, audience, and task title.

  2. 2

    Add two or more fal catalog endpoints. Use the advanced field for a private endpoint ID or ephemeral fal.run URL.

  3. 3

    Select compatible prompts and attach required reference media.

  4. 4

    Inspect each endpoint's current input contract and bind prompt, seed, references, variables, fixed values, or API defaults.

  5. 5

    Review the expanded items and exact request preview, then start generation.

  6. 6

    Watch the live task grid until every required candidate is complete or a failure has been resolved.

Endpoint identity and schema resolution

The selected endpoint ID remains the request target and the named comparison arm. GenMedia reads its current contract and may use a compatible alias only to locate the schema path used for validation. It does not redirect generation to that alias or guess input fields from a model name.

If a contract is missing, ambiguous, or incompatible with the selected prompts, task creation stops before generation. Fix the endpoint or mapping instead of forcing an uncertain request.

Import finished media

Use import for candidates produced by another provider, outputs generated before the task existed, rating tasks, and prompt or parameter treatment comparisons. Each item should include the shared prompt or case context, the candidate media, and enough labels to identify the arm after blind review ends.

GenMedia validates the document before it creates the task. Placeholder or incorrect URLs, malformed item groups, and missing candidates should be fixed at the source rather than repaired after voting begins.

Runs and extensions

A run records one generation or import batch. Adding prompts or endpoints to an existing task creates another run and keeps earlier items intact. Once evaluation assignments exist, changes that would reinterpret an earlier vote are blocked; create a new run or task version instead.

Minimal import shape

Two candidates, one comparable item

{
  "schema": "deltaeval-task-v1",
  "title": "Release candidate comparison",
  "task_type": "t2i",
  "evaluation_mode": "comparison",
  "visibility": "all",
  "criteria": ["prompt adherence", "visual quality"],
  "items": [{
    "prompt": "A product evaluation on a neutral background",
    "seed": 4102,
    "candidates": [
      { "endpoint": "production", "media": { "kind": "image", "url": "https://your-cdn.example/a.webp" } },
      { "endpoint": "candidate", "media": { "kind": "image", "url": "https://your-cdn.example/b.webp" } }
    ]
  }]
}

Replace placeholder URLs with media you control. See Imports and exports for the complete contract.