---

name: interactive-colorization description: "Guides agents using interactive-deep-colorization local-hints colorization APIs, GUI semantics, masks, color suggestions, and saved outputs." disable-model-invocation: true metadata: disco-role: operating license: MIT

interactive-colorization

Use this sub-skill when a task is about the repository's local-hints interactive colorization workflow: explaining the GUI, reproducing the notebook-style API flow, preparing user hint tensors, choosing PyTorch versus Caffe wrappers, interpreting recommended colors, or handling saved local-hints artifacts.

Route first

• Read references/api-reference.md for class signatures, backend wrapper behavior, CLI defaults, and PyTorch tensor shapes.
• Read references/workflows.md for GUI interaction semantics, a distilled notebook-style local-hints recipe, and saved-result handling.
• Read references/data-formats.md before creating or consuming input_ab, mask, Lab arrays, recommendation distributions, or saved .npy/.png artifacts.
• Read references/troubleshooting.md when imports, GUI launch, backend choice, tensor shape, gamut, color suggestions, or output saving behavior is confusing.

Boundaries

• For installation, dependency, display, backend setup, and model file acquisition questions, route to ../setup-and-models/SKILL.md.
• For global histogram transfer or reference-image workflows, route to ../global-histogram-transfer/SKILL.md.
• Training is out of scope for this repository skill; the repository points local-hints training to a separate external project.
• Do not treat Caffe, PyQt GUI execution, or downloaded model-weight inference as construction-verified here. Source/API facts and PyTorch architecture smoke behavior are verified; full native GUI/model runs need the setup/model sub-skill gate.

Safe bundled scripts

• Use scripts/inspect_cli_defaults.py to inspect distilled GUI CLI parser facts without importing PyQt, qdarkstyle, Caffe, or model weights.
• Use scripts/smoke_core_helpers.py with a repository checkout passed via --repo-root to check safe imports, Lab/gamut helpers, wrapper image preparation, and tiny PyTorch architecture forwards without loading model weights or GUI/Caffe modules.

Critical invariants to preserve

• The local-hints model consumes input_ab with shape 2 x Xd x Xd and input_mask with shape 1 x Xd x Xd before wrapper-specific normalization.
• Verified constructor and method signatures include ColorizeImageBase(Xd=256, Xfullres_max=10000), ColorizeImageTorch(Xd=256, maskcent=False), ColorizeImageTorchDist(Xd=256, maskcent=False), ColorizeImageCaffe(Xd=256), ColorizeImageCaffeDist(Xd=256), ColorizeImageCaffeGlobDist(Xd=256), ColorizeImageTorch.prep_net(self, gpu_id=None, path='', dist=False), SIGGRAPHGenerator(dist=False), and SIGGRAPHGenerator.forward(self, input_A, input_B, mask_B, maskcent=0).
• The GUI CLI has a parser quirk: --dist_model stores into dest='color_model', so there is no independent args.dist_model; both PyTorch wrapper initializations read args.color_model.