Release 1.2.1

[LightCyan01]

Summary

TrainKit 1.2.1 fixes NCNN model compatibility, supports processing individual images, makes resumable manifests opt-in, and standardizes Windows distribution on a ZIP-only release.

Changelog

Added

• Automatic NCNN input and output blob discovery through the Python bindings.
• Individual-image input support for captioning, tagging, upscaling, and renaming.
• Separate file and folder selectors in the desktop interface.
• Optional PFX signing in the Windows release workflow.

Changed

• Resumable manifests are now opt-in for normal runs and are still always produced for dry runs.
• NCNN blob names are optional overrides instead of required defaults.
• The standard Windows release is a ZIP archive; the installer has been removed.
• Release notes include extraction instructions and an unsigned-build warning when no signing certificate is configured.

Fixed

• NCNN models that do not use in0 and out0 no longer fail with an input-blob-not-found error.
• Matching NCNN .bin weights can be discovered automatically from the selected .param file.
• Batch errors now expose the first image-level failure when no manifest is written.
• Image preview and processing paths now accept either a single image or a directory.