Models and cache¶

AIsketcher does not bundle model weights. Guided Sample requires none; local generation uses pinned Hugging Face repositories after the user confirms a preset’s license, files, cache destination, and estimated download size.

Preset cards¶

Studio option	Versioned preset	Intended use	Approximate first download
Guided Sample	bundled fixture	learn the workflow	none
FLUX.2 Klein Edit — recommended	`flux2-klein-edit@1`	fast sketch rendering, photo restyling, and instruction edits on a 16 GB NVIDIA T4	about 16.2 GB
SDXL Canny Lite — legacy	`sdxl-canny-lite@1`	replay existing SDXL manifests and lower-memory edge conditioning	about 7.3 GB
SDXL Canny Quality — legacy	`sdxl-canny@1`	replay existing full-ControlNet SDXL recipes	about 9.4 GB
Korean→English helper	pinned `facebook/m2m100_418M` adapter	protect recognized design terms, then prepare model-facing English while preserving Korean input	about 1.9 GB when absent

FLUX.2 Klein Edit is the default for new live studies. It uses the uploaded image as a reference and does not use Canny ControlNet. The two SDXL presets remain available for explicit installation and replay when needed, but are labeled Legacy rather than presented as the modern default.

Sizes cover only the pinned, allow-listed files at the revisions below. Studio shows the model’s intended use, expected transfer, cache policy, device expectation, and license confirmation; plan_install() exposes the exact destinations for programmatic review before downloading.

The Studio preparation layer also shows the pinned Korean→English helper when it is missing. Review & prepare model explicitly confirms the selected image model and helper together. Leaving setup before pressing that button performs no network access.

Pinned repositories¶

The curated image presets and local translation helper use immutable revisions of:

Preset	Role	Repository	Revision
FLUX.2 Klein Edit	base/edit	`black-forest-labs/FLUX.2-klein-4B`	`e7b7dc27f91deacad38e78976d1f2b499d76a294`
FLUX.2 Klein Edit	decoder	`black-forest-labs/FLUX.2-small-decoder`	`a3efc24f613ef42d9428af62fdbd6f5fd8856c4a`
SDXL Canny legacy	base	`stabilityai/stable-diffusion-xl-base-1.0`	`462165984030d82259a11f4367a4eed129e94a7b`
SDXL Canny legacy	lite control	`diffusers/controlnet-canny-sdxl-1.0-small`	`edd85f64c5f87dfb6d73762949d9daca16389518`
SDXL Canny legacy	quality control	`diffusers/controlnet-canny-sdxl-1.0`	`eb115a19a10d14909256db740ed109532ab1483c`
Korean→English helper	translation	`facebook/m2m100_418M`	`55c2e61bbf05dfb8d7abccdc3fae6fc8512fd636`

Image-model loading is restricted to SafeTensors with remote code disabled. A mutable branch name is not sufficient for a replayable preset. FLUX.2 Klein’s pinned weights are Apache-2.0; the legacy SDXL components retain their pinned upstream OpenRAIL terms. The Korean→English helper is MIT-licensed and uses its pinned PyTorch state dictionary through the weights-only loader; its license is shown separately from the selected image model's license during preparation.

FLUX.2 Klein Edit resolves to at most 1024 × 1024, four steps, guidance 1.0, the Flow Match Euler scheduler, and reference-image control. The two SDXL presets resolve to 1024 × 1024, 30 steps, guidance 5.0, UniPC, and Canny ControlNet before an Advanced override. Structure strength resolves to 0.55 for Loose, 0.75 for Balanced, and 0.95 for Faithful. These values belong to versioned presets and may change only under a new preset version.

Cache behavior¶

Weights stay in AIsketcher’s platform user cache, outside the repository, wheel, and exported study. A manifest records repository and immutable revision, never an access token or the local cache’s absolute path. The packaged Studio keeps image-model directories below models/ and its pinned Korean helper below translation/ under the same resolved cache root. It does not depend on an unrelated process-wide Hugging Face cache, so a helper prepared through Studio remains available to that Studio in cache-only mode.

The cache root resolves in this order:

an explicit PresetManager(cache_dir=...) argument;
cache_dir loaded by the packaged Studio from its YAML configuration;
AISKETCHER_CACHE_DIR;
$XDG_CACHE_HOME/aisketcher when set;
~/Library/Caches/AIsketcher on macOS;
~/.cache/aisketcher on other supported platforms.

To keep Studio model files on another volume, set cache_dir in the per-user configuration created by aisketcher init:

schema_version: 1
cache_dir: "/Volumes/Models/AIsketcher"

For a direct Python integration or temporary shell override:

export AISKETCHER_CACHE_DIR=/Volumes/Models/AIsketcher

See Configuration for precedence and the complete schema.

Removing a cache is an explicit user operation. The Studio reports what will be removed and does not delete unrelated model caches.

The pinned M2M100 Korean→English revision publishes seven runtime files: weights, model and generation configuration, and tokenizer vocabulary/config. AIsketcher verifies the reviewed size and SHA-256 of every one before loading the local snapshot, and disables Transformers’ optional background SafeTensors conversion helper during that load. This prevents an independently moving refs/pr/* conversion artifact from being fetched and keeps the reviewed revision authoritative. The exact allow-listed first transfer is about 1.9 GB. Any pre-existing process policy for that Transformers setting is restored immediately afterward.

Before tokenization, a deterministic glossary replaces recognized Korean visual-design terms with their reviewed English production vocabulary. Studio then runs Korean→English translation locally and records the exact source text, prepared English, helper ID, and immutable revision as separate prompt provenance. The glossary is intentionally bounded and does not imply that every Korean phrase will translate perfectly.

First download, Stop, and retry¶

Selecting an uncached model does not immediately start network access. Review the repositories, immutable revisions, expected transfer, destination, license, and device guidance, then confirm the separate download action. allow_downloads: true permits this flow but is never consent by itself.

During image-model preparation, Stop is checked between selected download groups and between streamed SHA-256 chunks. It never turns a partial or unverified tensor into a usable cache. Stopping a read-only integrity pass leaves the existing files in place for a later retry. A failed fresh download removes only that incomplete managed destination; it never removes another installed preset or follows a symlink outside the managed cache.

The Korean helper uses the pinned Transformers cache below translation/ rather than AIsketcher image-model markers. Its cooperative Stop check occurs at every reviewed file boundary, during streamed hashing, and between the tokenizer and model loads. A later retry can reuse files already present for the same immutable revision, but the image-model marker-and-cleanup guarantee above does not apply to that cache format.

The completion marker uses the versioned aisketcher-model-cache schema and records the repository, immutable revision, allowlist, hash policy, reviewed artifact sizes and SHA-256 values, and the verified file-stat fingerprint. The marker is provenance, not an authentication token: a fresh Studio process does not trust it by itself. Policy v2 streams every exact curated runtime file — configuration, scheduler, tokenizer, index, and LFS weight payloads — through the reviewed SHA-256 policy before the backend may load it. A process-local receipt then permits fast reuse only while the marker and every verified file's size, modification/change time, device and inode remain unchanged. Legacy markers are fully verified and upgraded; a forged marker, configuration edit, or same-size tamper is rejected.

Opening Studio uses a non-verifying display plan so a multi-gigabyte hash pass cannot hide the page behind a connection delay. Until Review & prepare model finishes, an existing cache may therefore appear as Not yet verified · download if absent and the displayed transfer is a conservative maximum, not proof that it will be downloaded. On the validated T4, checking the roughly 16.2 GB FLUX cache took up to about one minute with zero network transfer. The same process reuses its receipt immediately afterward.

After a successful first download, later Studio sessions load the same pinned files from the configured local cache; they do not redownload them merely because the browser was refreshed.

Refreshing is not cancellation. If the same browser session reconnects while a job is running or stopping, Studio restores that status and the Stop control. If the temporary server itself ended, the connection-recovery layer directs the user to the latest Studio address.

Install from Python¶

Planning is read-only. Installation requires a separate confirmation after the application has displayed the plan.

from aisketcher import PresetManager

manager = PresetManager()
plan = manager.plan_install("flux2-klein-edit@1")

print(plan.license_notice)
print(plan.estimated_bytes, plan.download_bytes)
for item in plan.items:
    print(item.repo_id, item.revision, item.destination, item.installed)

# Call only after the user reviews the information above.
result = manager.install("flux2-klein-edit@1", confirm=True)

PresetManager(allow_downloads=False) is the required setting for a hosted or read-only environment where an administrator, rather than a web user, manages the model cache.

Custom repositories¶

Arbitrary repository URLs, single checkpoint files, custom Python pipelines, pickle weights, and trust_remote_code are outside the curated Studio path. Implement a backend in Python when a controlled deployment needs another model; do not expose a public URL field that turns the Studio into a remote-code loader.