---

name: refinement-and-export description: "Operate Make-It-3D refinement, test rendering, multi-view generation, point-cloud/mesh export, and output troubleshooting after coarse training." metadata: disco-role: operating disable-model-invocation: true license: NO_LICENSE

Refinement and Export

Use this sub-skill after a coarse Make-It-3D workspace exists, or when a user asks about --refine, --test, videos, checkpoints, multi-view images, point clouds, OBJ mesh export, or export-specific dependencies.

What This Sub-Skill Covers

• Refinement command construction and the source control-flow quirk around --final --refine.
• Test rendering from checkpoints and output locations.
• Mesh export with --save_mesh and its xatlas/nvdiffrast requirements.
• Point-cloud/multi-view/refine utilities and PyTorch3D/Open3D dependencies.
• Output validation and troubleshooting after partial/interrupted runs.

Setup/input issues belong to environment-and-inputs. Coarse command generation belongs to coarse-training.

Required Reads and Scripts

• Read references/workflow.md for refine/test/export command flow.
• Read references/outputs-and-export.md for workspace files, videos, OBJ/texture outputs, and dependency ownership.
• Read references/troubleshooting.md for common refine/export failures.
• Run scripts/build_refine_export_commands.py to generate copyable commands.

Command Pattern

python /path/to/skill/sub-skills/refinement-and-export/scripts/build_refine_export_commands.py \
  --workspace NAME --ref-path REF_ALPHA.png --text "object prompt" --save-mesh

The helper emits:

• a refine command using --final --refine to satisfy the inspected source nesting;
• a test render command with --test;
• an optional mesh export command with --test --save_mesh.

Safety and Verification

• Do not start refinement unless the coarse workspace has checkpoints and the user accepts additional GPU time.
• Do not promise OBJ texture export until xatlas and nvdiffrast import in the runtime environment.
• Treat PyTorch3D/Open3D/contextual-loss import errors as refinement/export blockers, not generic Python issues.
• Check generated outputs before deleting or rerunning a workspace. Interrupted training can still leave useful checkpoints.