Skip to content

AIsketcher 0.4.0

Version 0.4.0 is a first-use and model-trust release. The package still preserves the same study and replay contracts, but a new user can now see those contracts before committing to a local model runtime.

Install

Layer Command Includes
Core python -m pip install "aisketcher==0.4.0" SDK, export/replay, bundled tour
Studio python -m pip install "aisketcher[demo]==0.4.0" core plus Gradio Studio
Local python -m pip install "aisketcher[local]==0.4.0" core plus local model runtimes
Full local python -m pip install "aisketcher[local,demo]==0.4.0" local generation and Studio

The quickest proof is:

aisketcher try

It opens the bundled, hash-verified study on localhost with no model or external network request.

Honest model roles

The former Simple-mode Auto choice routed both sketches and photos to the same FLUX.2 Klein preset. It did not inspect the input, hardware, or intended constraint. Version 0.4 retains auto as a compatibility alias but removes it from the normal selector and displays the concrete role instead.

  • Fast Edit · FLUX.2 Klein is for photo restyling, flexible sketch interpretation, and instruction edits. It is not a Canny/ControlNet route.
  • Structure Lock · SDXL Canny Lite is the lower-memory legacy line-control fallback.
  • Structure Lock+ · SDXL Canny is the full legacy ControlNet path.

Z-Image Turbo + Union 2.1 Lite and Qwen Image Edit stay in the curated research registry until they pass the published multi-input, four-seed benchmark. A large or popular checkpoint does not become the default without actual AIsketcher output evidence.

First-use protection

The packaged Studio supplies its configured device to the controller. Before the model installer writes or downloads files, it checks:

  • whether live CPU generation is unsupported;
  • whether FLUX.2 was selected for an unvalidated MPS environment;
  • whether CUDA is available when required;
  • whether known VRAM is below the reviewed minimum; and
  • whether the model cache has enough free space for remaining files and working headroom.

These checks do not change the lower-level PresetManager API used to prepare a cache for another machine. They protect the interactive local product path, where downloading a model that cannot run is almost always a mistake.

English Studio preparation now downloads only the selected image model. The pinned Korean→English helper appears and is prepared in the Korean workflow, removing an unnecessary 1.9 GB transfer from the English first run.

Simple mode now starts with one quick preview instead of an implicit four-image job. Users can choose four directions for comparison or eight for a wider seed search, with the time tradeoff visible before generation begins.

Discoverability and feedback

The repository README now leads with an actual Studio capture and the model-free command. Package keywords describe the current model-independent design-study product. Structured issue forms collect reproducible bugs and candidate-model evidence without encouraging screenshots that omit seeds, runtime, or permissions.

GitHub Pages now builds and deploys on each push to main; the same strict documentation tests still run in CI and in the Pages build. The release workflow remains versioned and publishes to PyPI through Trusted Publishing.

Compatibility

  • Existing auto values continue to resolve to flux2-klein-edit@1.
  • Existing versioned presets and manifests remain valid.
  • The base package remains free of Torch, Diffusers, and Gradio dependencies.
  • The new aisketcher try command uses only the standard library and packaged assets.

License boundary

Source code and documentation text are MIT licensed. Bundled drawings, screenshots, generated outputs, and other artwork remain outside the MIT grant; read the artwork notice before reuse.