Dependencies and limits#

Rendering can involve Keras/TensorFlow plus Pillow and optional drawing/font backends. Missing optional packages should be diagnosed at the operation that needs them rather than turning package import into an eager dependency check.

The maintenance snapshot currently records open findings around graph-shape handling, ellipsizing, optional backends, default/deprecation behavior and test matrix coverage. Treat representative gallery output as examples, not proof that every custom layer/framework version is supported.

Large models can create very large images or unreadable graphs. Use filtering, scaling and focused sub-model diagrams when a single “whole network” picture stops communicating useful structure.