AIsketcher 0.2.0

Publication source of truth

This page describes the 0.2.0 artifacts. A source checkout or tag is not proof of PyPI publication; verify the version on the PyPI project page.

A design study instead of a one-shot image

AIsketcher 0.2.0 turns one sketch into a traceable family of directions. A run now records preparation, resolved generation controls, every candidate seed, a human-selected parent, controlled variations, file hashes, and runtime drift. The backend is replaceable; the study contract is stable across backends.

The Pocket Kingdom sample is the canonical example. It includes the anonymous source, exact prepared control, four real local scout directions, the selected parent, four structure-locked variations, final human selection, and both manifests.

The base wheel contains the Studio code, verified Guided Sample, and console entry point. The demo extra adds the Gradio runtime needed to launch them. aisketcher init writes an owner-only, schema-versioned YAML file; aisketcher studio applies its language, preset, device, output, seed, cache, and download-policy defaults without requiring a repository checkout.

Upgrade impact

The new package is intentionally not a drop-in replacement for 0.0.x:

• change application imports from AIsketcher to aisketcher;
• replace one-shot img2img(...) calls with the complete SDK workflow;
• remove AWS credentials and cloud translation options from configuration;
• choose a backend and versioned preset explicitly;
• store exported studies rather than isolated result files.

The deprecated compatibility facade provides a short transition for a caller-owned pipeline. Follow the migration guide before 0.3.0.

Install profiles

Profile	Command	Includes
Core	python -m pip install "aisketcher==0.2.0"	preparation, studies, export, replay
Studio	python -m pip install "aisketcher[demo]==0.2.0"	core plus local Gradio interface
Local	python -m pip install "aisketcher[local]==0.2.0"	core plus Diffusers runtime
Full local	python -m pip install "aisketcher[local,demo]==0.2.0"	local generation and Studio

These commands require PyPI to list 0.2.0. For a local release candidate, use the reviewed wheel, an editable checkout, or a pinned source revision.

Model weights are never inside the wheel. Curated presets expose the pinned repositories, revisions, cache destinations, estimated bytes, and license notice before an explicit confirmed installation.

Runtime support

• Python 3.10 through 3.14;
• CUDA for supported NVIDIA environments;
• Apple Silicon MPS as an experimental sequential-generation path;
• no live CPU generation in the v0.2 Studio;
• deterministic FakeBackend for tests and integration examples.

Pixel identity is bounded by the complete runtime, not only a seed. Strict replay checks recorded resources and reports drift; it does not promise that different accelerators produce identical pixels.

Publication checklist

The release is ready to publish only after all of the following are true:

1. core, Studio, documentation, distribution, and wheel-install smoke tests pass from a clean checkout;
2. the version in the tag, wheel metadata, package, and release notes is exactly 0.2.0;
3. the GitHub release is created from v0.2.0;
4. trusted publishing sends that exact artifact to PyPI;
5. a fresh environment can install the PyPI wheel and open Guided Sample;
6. the documentation site is deployed from the same revision;
7. the PyPI project description and public links render correctly.
8. docs/assets/aisketcher-social-preview-github.jpg is uploaded manually in GitHub Settings → Social preview; the repository file alone does not change that setting.

Publication is a separate, explicit maintainer action. Building a wheel locally does not satisfy this checklist.

License boundary

Source code and documentation text are MIT licensed. Pocket Kingdom drawings, generated derivatives, and other visual assets are excluded. Read the artwork notice before reusing any image.