One sketch. A traceable family of directions.

AIsketcher is a model-agnostic Python SDK for structured visual exploration, controlled variation, and reproducible creative handoff.

It treats image generation as a design study:

prepare → explore → pick → vary → export → replay

The package keeps this workflow separate from any particular checkpoint or cloud provider. A backend resolves a model-independent intent into capabilities it can actually support; the resulting recipe, seeds, artifacts, and lineage are then recorded together.

Get started 한국어 빠른 시작 See the Studio

Pocket Kingdom hero concept — marketing artwork, not an SDK execution claim. Open the canonical sample for the real local inputs, candidates, seeds, selections, and replay manifests.

python -m pip install "aisketcher[demo]==0.3.0" && aisketcher init && aisketcher studio

The install identifier, Python import, and CLI are all lowercase aisketcher. This launches Studio. Select Try the guided sample to open the bundled fixture without downloading a model. Its results are read-only: Refine this direction opens the model-preparation layer rather than an error, and Keep exploring the sample closes it without changing the fixture. Run aisketcher init only for the first launch; omit it when settings already exist. The configuration reference explains the YAML and project overrides.

Actual local English Studio with the bundled HPO Guided Sample fixed to seed 6764547109648557242 and pinned historical sdxl-canny-lite@1 provenance. This legacy fixture is not the live default. Select the image to inspect the full-size interface.

The bundled, privacy-reviewed fixture displays its prompt, profile, and structure directly from the authenticated manifest. Twelve new candidates were reviewed in four bounded HPO rounds before the selected direction was captured. It opens without model weights, network access, or an image upload. Pocket Kingdom remains a separate documentation-only canonical lineage example. New live studies use Auto, which selects the recommended T4-validated FLUX.2 Klein Edit path. SDXL Canny remains available for legacy replay or intentional edge-conditioned work.

Built for design decisions

Stage	Design question	Recorded evidence
Prepare	Is the sketch usable as structure?	normalized source, control, diagnostics
Explore	Which directions are worth seeing?	intent, resolved recipe, candidate seeds
Pick	Which candidate becomes the parent?	selection and technical badges
Vary	What may change and what stays locked?	parent ID, strength, constraint locks
Export	Can another person inspect the work?	images, contact sheet, manifest, hashes
Replay	Can the run be reconstructed honestly?	model revision, drift report, runtime

What it does not promise

AIsketcher does not claim that one seed is universally “best,” that a model is deterministic across every hardware stack, or that a technical score determines the most beautiful result. Recommendation badges describe observable structure and diversity properties; creative selection remains with the designer.

Try without a model

Studio Guided Sample uses a reviewed, hash-verified local fixture. It does not download weights or require a GPU. The bundled v0.3 fixture is the privacy-reviewed HPO hero study with selected seed 6764547109648557242. Pocket Kingdom is a separate documentation-only canonical lineage example that shows input, seeds, results, variation, and evidence together. Preparing a live model is a separate, confirmed action that also shows the pinned Korean→English helper when it is missing; closing the layer performs no download.

Artwork has a separate license

Code and documentation text are MIT licensed. Drawings and sample images are not. Read the project’s artwork notice before using a visual asset.

Release status

Version 0.3.0 is available from PyPI. Its package description is the README embedded in the release artifact, so publishing the tagged GitHub Release updates the PyPI introduction together with the wheel and source archive. Editing main or an existing Release does not rewrite an immutable PyPI version. The documentation site is deployed separately by the current manual-only Pages workflow. See the changelog and 0.3.0 release notes for the current boundary.