Skip to content
GenMediaOpen workspace
Browse documentationInputs and references

Keep shared inputs visible and traceable

Prompts, seeds, reference assets, fixed fields, and sweep values define the case a reviewer is judging. GenMedia stores them with the item so the generation can be reproduced and the review can expose the source material when it matters.

Prompt Library and generated Datasets

Prompt Library is the reusable prompt bank. Each entry stores reusable text, modality, workflow requirements, references, and tags; prompt groups keep a curated selection together.

A Dataset contains the examples generated from those prompts. For example, 30 prompts generated with three models produce 90 examples. Publishing those outputs creates an immutable Dataset version that can be inspected and bound to an Evaluation Task. Regenerating the same definition as a later version makes model progress comparable over time.

Tags describe the content and use case. They support later slices and diagnostics, but they do not decide which candidate won.

Prompt matching is exact and fail-closed

GenMedia never matches prompts by title similarity, filename, or array position. Generation and task-extension requests carry canonical Prompt Library IDs. The server loads those exact non-archived rows, restores the requested order with an ID map, and rejects duplicates, missing IDs, archived prompts, or workflow-incompatible prompts before it creates jobs.

Finished-media imports may attach an explicit prompt_ref to an item. That reference is resolved by exact ID and the task stores both the canonical prompt ID and the task-time prompt snapshot. An omitted prompt_ref remains intentionally task-local; GenMedia does not invent lineage from similar text.

Reference media

  • Image references

    Use for image-to-image, image-to-video, product, style, and identity evaluations.

  • Video references

    Use when motion, timing, or source footage is part of the case.

  • Audio references

    Use for speech, music, sound, or audiovisual alignment evaluations.

  • Reviewer inspection

    Reference-dependent evaluations expose an Inputs control so reviewers can inspect the source without revealing model identity.

Variables and sweeps

A prompt can include declared tokens such as @subject or @environment. The planner combines their values with endpoint field bindings and shows the resulting item and job count before generation.

Use an explicit value list or numeric range for a sweep. A sweep expands the dataset. It should not be used to hide several experimental arms inside one item.

Fixed values and API defaults

Set a fixed value when it is part of the Evaluation Task control. Leave a field on its API default only when the endpoint contract marks it optional and the default is acceptable for every candidate arm. Record the important control choices in the task notes so another operator can understand the evaluation later.

Prompt contract

Declare every token before materialization

Prompt
"A @subject moving through @environment"

Variables
subject: ["cyclist", "runner"]
environment: ["city at night", "coastal road"]

Result
2 x 2 = 4 controlled prompt cases per endpoint