nvidia-omniverse/usd-content-agents/content-texture-candidate

Texture Candidate Generation

Generate or regenerate candidate artifacts for exact plan unit IDs. Embedded generation is non-mutating: it cannot publish or replace the canonical asset.

When to Use

Use after provider-free preparation and an outer-authored plan selects generate. For refinement, use only exact outer revise unit IDs.

Limitations

• A remote Texture Agent may be a leaf backend, never the workflow controller.
• Provider generation and deterministic supplied-image apply are mutually

exclusive. Neither mode falls back to the other.

• Reviewed unit artifacts are immutable preserved inputs during refinement.
• Generation cannot mutate geometry, physics, collision, or unrelated material state.
• A leaf backend cannot change the outer-accepted targets, appearance,

generator inputs, preservation constraints, or acceptance criteria.

Prerequisites

Load texture_preparation.json, the exact outer plan, and any explicitly requested advisory proposal. A proposal is not required for other generator leaves.

Instructions

1. Verify generate=requested; reject not_requested rather than silently

generating.

1. Bind the outer plan's exact targets, semantic intent, reference identities,

and generator inputs. Do not replace them with advisory proposal values.

1. Select exactly one execution mode. For provider_generate, select the

provider/capability explicitly and record its provenance. For apply_provided, bind one ordered outer_generated_candidate albedo PNG to each exact unit, including its producer capability and invocation provenance.

1. Invoke only the selected leaf. apply_provided performs deterministic USD

authoring and must not call Texture Agent, simple image generation, a model, or another provider.

1. Persist the generation result, candidate asset, unit artifacts, and digests

before deterministic evidence collection.

1. Preserve reviewed artifacts byte-for-byte during bounded revision.
2. Stop after generation. Do not render, critique, publish, or update the

canonical asset.

Command Reference

Run content-workflow-cli texture generate --preparation ... --outer-plan ... with an explicitly exposed generator provider, or run content-workflow-cli texture apply-provided --preparation ... --outer-plan ... for outer-created images. The bundled Texture-service adapter opens an execution session from the deterministic scope and exact outer generator inputs. Pass the optional --provider-proposal to either command only when propose=requested; it remains advisory and never drives mutation. The remote service adapter requires a byte-self-contained source and fails closed when the prepared USD has an external dependency closure it cannot upload. Typed Python adapters can implement TextureGeneratorLeaf, while the public CLI supports only its concrete named adapters. Do not claim arbitrary adapters are callable until they are explicitly exposed.

Common Workflows

After refine, route immediately to content-texture-quality; never replay the initial plan or reviewed candidates. Resume must reuse the exact persisted candidate/result and cannot call the generator again.

Output Format

Retain texture_generation.json, the exact outer-plan and optional-proposal bindings, execution mode, generator/apply capability, generator inputs, candidate binding, per-unit artifacts, references, outer-provided image bindings and producer provenance, and operation statuses.

Troubleshooting

If the provider returns unknown, missing, duplicated, or reordered unit IDs, or if supplied images are missing, duplicated, stale, symlinked, reordered, scope-mismatched, or not exact-size PNG albedo artifacts, stop. Do not translate display names into plan-unit identities or treat references as candidates.