Skip to content

Changelog

This page tracks user-visible package changes. PyPI is the source of truth for whether a listed version is publicly available.

0.3.0

Version 0.3.0 modernizes the default local model path and makes Studio’s download, refinement, cancellation, and recovery boundaries explicit.

Added

  • a T4-validated FLUX.2 Klein Edit backend and curated Auto route for new sketch-to-design, photo-led, and instruction-edit studies;
  • an explained model choice in Simple, including cache state, pinned revisions, expected transfer, device guidance, and upstream license notices;
  • an explicit, pinned local Korean→English helper that preserves the Korean original and records the prepared model prompt and translator revision, without Transformers fetching an unpinned background conversion PR;
  • cooperative Stop behavior for generation and model/helper preparation, same-session job reconnection, and a recovery layer for an ended temporary Studio server;
  • a refinement instruction composer and a model-preparation layer for attempts to refine the read-only Guided Sample.

Changed

  • flux2-klein-edit@1 is the recommended/default live preset; SDXL Canny Lite and Quality remain versioned Legacy choices for manifest replay and intentional edge-conditioned work;
  • the four Simple direction cards are larger and avoid an inner gallery scrollbar, while progress distinguishes elapsed time from an estimate;
  • model preparation requires an explicit confirmation before downloading the selected image model or the Korean helper. Cancelling the layer performs no network access, and a retry reuses complete pinned cache entries;
  • packaged Studio resolves the Korean helper below the same configured AIsketcher cache root as image models, so cache-only reuse does not depend on an unrelated global Hugging Face cache;
  • FLUX.2 variation levels are applied through explicit deterministic edit instructions and recorded as an approximation, while model downloads observe Stop between curated file groups;
  • release and deployment actions are pinned to reviewed full commit SHAs;
  • current install examples use the lowercase aisketcher identifier and pin version 0.3.0;
  • the README embedded in each release artifact remains the PyPI project description. Publishing the matching GitHub Release publishes that immutable description and package through Trusted Publishing.

Removed

  • the deprecated uppercase AIsketcher.img2img facade;
  • the deprecated aisketcher.modelPipe compatibility module and uppercase top-level module packaging.

The historical Pocket Kingdom and heritage manifests keep their recorded SDXL recipes, seeds, hashes, and AIsketcher 0.2 runtime provenance. See the detailed 0.3.0 release notes.

0.2.1

Version 0.2.1 makes the modern v0.2 package the default public PyPI experience and aligns its presentation with the reviewed Studio workflow.

Changed

  • installation examples now use version-pinned PyPI commands for the core, Studio, local-generation, and combined optional dependency profiles;
  • the README carried into PyPI metadata describes the current model-agnostic SDK rather than the historical AWS workflow;
  • the English Studio documentation now leads with an actual local capture of the HPO-selected heritage direction: seed 6764547109648557242 with pinned preset sdxl-canny-lite@1;
  • publishing a GitHub Release for a matching v<version> tag now starts the guarded PyPI workflow automatically. The job verifies the tag and package version, runs release tests and distribution scans, smoke-tests the wheel, and publishes through OIDC Trusted Publishing without a PyPI API token.

See the detailed 0.2.1 release notes.

0.2.0

The v0.2 line rebuilds AIsketcher around a model-independent design study:

prepare → explore → pick → vary → export → replay

Added

  • normalized sketch preparation with Canny diagnostics;
  • explicit seed plans and ordered 1, 4, or 8 candidate studies;
  • human selection, parent-child variation lineage, constraint locks, and technical recommendation badges;
  • portable exports with manifests, file hashes, runtime evidence, strict replay, and compatible replay reports;
  • curated, immutable SDXL Canny presets with review-before-download plans;
  • a Simple/Advanced Gradio Studio and a hash-verified, model-free Guided Sample;
  • a packaged aisketcher init / aisketcher studio CLI and versioned YAML settings ledger;
  • the Pocket Kingdom canonical scout and variation study with complete replay evidence.

Changed

  • new code imports the lowercase aisketcher package;
  • generation is separated from the SDK through a backend capability protocol;
  • local model runtimes and Studio are optional installation layers;
  • the Studio and its reviewed Guided Sample now ship in the wheel instead of requiring a repository checkout;
  • seeds are presented as reproducibility coordinates, not universal quality scores.

Removed

  • AWS-specific translation, credential, and cloud execution paths;
  • arbitrary model URL loading and remote-code execution in the Studio;
  • implicit downloads during import or ordinary SDK construction.

Security and privacy

  • model presets pin repository revisions, allow-list fp16 components, require SafeTensors, and disable remote code;
  • exports omit tokens, original upload names, absolute cache paths, and image metadata;
  • the Studio binds to loopback, disables public share links, limits uploads, and serializes generation by default.

Compatibility

The AIsketcher.img2img facade remains deprecated for the 0.2 line and requires a caller-provided compatible pipeline. It is scheduled for removal in 0.3.0. Cloud and credential arguments fail with a migration error instead of being accepted or logged.

See the detailed 0.2.0 release notes and 0.0.x migration guide.