Personal Image Model Studio

A local-first image-generation workbench by Jason Stys, with real private accounts, actual custom diffusion training, structured prompts and auditable datasets. Python handles learning and the local service; TypeScript/HTML/CSS provide the responsive interface; SQL indexes account-scoped records.

Download and open

Get the checksummed CPU application for your OS from Releases. Extract and open PersonalImageModelStudio (.exe on Windows). It launches a private website in your normal browser, with no Python/npm setup for native-model training or generation. Keep the application process running. This is an unsigned portable application, not a signed installer or public hosted service.

Create an alias/password, save the once-shown login recovery code and choose a separate vault phrase. Private files live outside Git and the installation. Locked workspaces use authenticated encryption; active ML needs temporary plaintext on your own device. OS administrators/compromised devices can access active data; crashes can leave protected recovery files. No telemetry, ad SDK, default external AI calls, automatic model downloads or developer account-access backdoor is included. Forgotten vault phrases cannot be recovered.

Private installation/help · Account security and limits · New validation evidence · Jason's contact

It does not generate placeholder images when no model exists. Train a fresh compact native model or explicitly register an authorized, compatible local classic Stable Diffusion export. No third-party artwork or pretrained weights are distributed here.

Rows: circle, square, triangle. Columns: red, green, blue. These samples came from real learned weights after 6,000 optimizer updates on owned procedural shapes, using the reference Torch 2.14 CPU environment—not a procedural renderer at inference time. The small model still produces rough edges and is not a general-purpose or photorealistic generator. See validation evidence, including exact runtime versions, separately retained earlier GPU evidence and limitations.

Features

• Independent scene, style, tag, negative and subject fields; explicit (phrase:1.4) weights; optional experimental native subject boxes.
• Real from-scratch diffusion: learned caption vocabulary/embeddings, small UNet, held-out denoising evaluation, EMA weights, safe checkpoint reload and AdamW/RNG resume.
• Text generation, image-to-image editing, white-edit/black-preserve masks and centered aspect-ratio expansion. Known pixels are overlaid exactly at output; semantic quality remains model-dependent.
• Directory/ZIP ingestion with caption and metadata sidecars, provenance, pixel deduplication, deterministic train/validation splitting and physically separate AI-origin quarantine metadata.
• Classic single-encoder Stable Diffusion safe-tensor imports and real attention-only LoRA training. Supported families are explicit; arbitrary executable models are rejected.
• Human feedback, explicit revisions and consented feedback ZIP exports. Generated pixels keep their AI-origin labels; ordinary training excludes them. No hidden retraining or automatic score inflation.
• Reviewed tag glossary, transparent local phrase rules, private run reports, bounded job queue, cancellation and interrupted-run recovery.
• Prioritized DeviantArt OAuth/PKCE connector with read-only operations for your own authorized gallery. Provider token permissions may be broader, including Sta.sh management: review the actual consent screen. Passwords stay in the provider browser; OAuth tokens stay in memory. Live account access requires your registered public client ID and authorization.
• Account login/recovery/deletion, account-scoped file pickers, encrypted idle snapshots, lock/sign-out, human-idle timeout and built-in help. Browser mode cannot select executable runtimes or read arbitrary PC paths.
• Responsive browser UI plus portable CPU packages built and smoke-tested on Windows amd64, macOS arm64 and Linux x86_64 before publication. Mobile viewport coverage is not a native iOS/Android installer or public mobile host.

Run locally

Python 3.12 and Node 24 are the reference toolchains. Choose an isolated environment; ML dependencies are substantial. CPU works without Docker. For a compatible GPU build, consult the official PyTorch installer and verify the version/driver combination rather than silently replacing system drivers.

python -m venv .venv
# Activate .venv using your operating system's normal activation command.
python -m pip install -e ".[dev,ml,desktop]"
npm ci
npm run build
python -m studio.private_cli

The launcher opens a token-free loopback URL at an available port. Sign in, then unlock the separate vault. Source mode with full ML extras preserves compatible classic SD/LoRA; the bundled CPU application supports native models only. python scripts/build_native.py builds that CPU package after the UI build. No model weights or GPU drivers are bundled.

Legacy developer mode: python -m studio.cli --port 8016 / --desktop and scripts/build_desktop.py remain for historical reproducibility. They use a single-user bearer-token workspace, not the new encrypted account mode. Do not expose either launcher to a public network. Existing legacy data is preserved and not silently migrated.

Installation and packaging · Walkthrough · DeviantArt setup

For a small real-artwork training example without an account, use the optional licensed sketch sample. The actual app test trained a fresh 6,000-step model on 384 sketches and verified generation/seed replay; measurements also record its rough scribble-like outputs and weak category fidelity. Downloads require explicit opt-in; existing synthetic examples remain available.

Live DeviantArt authorization and identity lookup were verified. The approved own-gallery sample returned no supported images, so account-artwork training remains unvalidated until eligible permitted images are supplied. The real synthetic custom-model evidence above is separate; see the live validation record.

Reproduce validation

python -m pytest tests/test_core.py tests/test_deviantart.py tests/test_quickdraw.py tests/test_jobs.py
python -m pytest tests/test_ml.py
python -m pytest tests/test_private.py tests/test_private_ml.py
python -m ruff check studio scripts tests
python -m ruff format --check studio scripts tests
npm run build
node scripts/code_index.mjs --check
npm run test:browser
npm run test:private
# RUN_ML_BROWSER=1 selects the separate real browser-to-ML-worker test.
python scripts/train_demo.py --root artifacts/my-demo --steps 6000 --device auto

Demo outputs/checkpoints are local and ignored by Git. Use a new output directory; existing models are never overwritten. CI repeats contract/account tests on three OSes/two Python versions, real CPU training/resume/LoRA and browser-worker tests, four browser profiles, dependency audits and legacy Windows packaging. A separate release gate requires green exact-source validation and real training/generation/encrypted-reopen tests of all three packaged executables. CI artifact uploads are opt-in (PIMS_UPLOAD_CI_ARTIFACTS=true); downloads use Releases rather than Actions storage. CI is not a GPU quality evaluation or proof of compatibility with every device/model.

Layout and file responsibilities

File/group	Responsibility
studio/contracts.py	Closed bounded input schemas and cross-field validation.
studio/api.py	Preserved internal ML API and legacy single-user routes, provenance and artifacts.
studio/cli.py	Legacy browser/native launch and per-process bearer session.
studio/private_cli.py, private_api.py	Default account launcher, loopback/CSRF boundary, account routes, uploads and scoped ML proxy.
studio/accounts.py	Scrypt passwords, session/recovery digests, throttles and versioned key configuration.
studio/private_workspace.py, vault.py, permissions.py	Exclusive local engine, streaming encrypted archive, protected roots, lock/recovery lifecycle.
studio/prompts.py	Numeric weight syntax, independent sections and phrase policy.
studio/files.py	Atomic JSON, hashes, confinement and metadata-stripping raster decode.
studio/datasets.py	Safe local/ZIP import, deduplication, splits, AI exclusion and owned fixtures.
studio/model.py	Fresh word vocabulary, weighted text/time conditioning and compact denoiser.
studio/training.py	Actual native optimization, held-out metrics, EMA and continuation.
studio/checkpoints.py	Fixed-architecture safe-tensor manifests, integrity and optimizer recovery.
studio/native.py	Learned DDIM sampling, regions and source-preserving edits.
studio/formats.py	Fixed pretrained classes, file budgets and base-model fingerprints.
studio/pretrained.py	Offline classic SD generation and real LoRA adapter training.
studio/jobs.py, worker.py	Serialized bounded subprocess lifecycle and trusted dispatch.
studio/store.py	Parameterized short-lived SQLite connections and quarantine database.
studio/feedback.py	Explicit latest-consent AI-labelled feedback export.
studio/deviantart.py	Own-gallery OAuth/PKCE, memory credentials, bounded metadata/raster downloads.
studio/__init__.py	Package marker.
web/main.ts, portal.ts, workspace.ts, api.ts, types.ts, style.css	Fail-closed mode selection, login/key derivation/help, ML UI, cookie transport, contracts and responsive styling.
index.html, tsconfig.json, package*.json, pyproject.toml	Entry document, strict compilation, locked browser dependencies and Python extras.
scripts/train_demo.py	Real fresh training and learned sample contact sheet.
scripts/build_desktop.py, desktop_entry.py	Native portable shell packaging/entry.
scripts/build_native.py, native_entry.py, smoke_native.py	Checksummed CPU executable, embedded worker entry and actual packaged account/ML validation.
scripts/private_browser_server.py, playwright.private.config.ts, tests/private*.spec.ts	Disposable profile services, real account browsers and genuine training workflows.
tests/test_private.py, test_private_ml.py	Ownership/authentication, boundary/crypto/corruption/recovery and actual encrypted ML lifecycle tests.
scripts/code_index.mjs, python_index.py	Reproducible exact declaration/variable line headers and detailed code map.
tests/test_core.py, test_deviantart.py, test_jobs.py, test_ml.py	Core/privacy/property, mocked provider-contract, lifecycle and real ML regression tests.
scripts/prepare_quickdraw.py, evaluate_quickdraw.py, tests/test_quickdraw.py	Opt-in licensed sketch sampling, genuine checkpoint/inference evaluation and offline bounded-input tests.
tests/browser*.spec.ts, playwright.config.ts	Real browser accessibility/portability and separate real ML-worker workflow.
.github/workflows/validate.yml, release-native.yml	Pinned validation/audits and exact-source three-platform draft-to-published release gates.
requirements-cpu-audit.txt	Upstream advisory lookup for pinned official CPU-wheel releases; not an installer.
scripts/publish_evidence.py, .gitattributes	Guarded synthetic-only public evidence export and stable text line endings.
docs/	Architecture, requirements traceability, threats, complexity, research, operating instructions and validation reports.
.gitignore, LICENSE, SECURITY.md	Private-artifact exclusions, software license and safe vulnerability reporting.

Detailed source map includes each declaration, argument/local variable, actual line and source-derived purpose. Header indexes are generated and checked in CI so line references cannot silently drift.

Scope and responsible use

Use only data/models you own or are permitted to process and train on. Account access is not a copyright/license grant. AI-origin detection uses explicit labels—not an infallible visual detector. A human checks captions, rights, image quality and safety. Accounts isolate local records, but only one account may unlock the engine at a time. This is not hardened multi-tenant hosting and must not be exposed to a public network. An archive download alone lacks account/key settings: cold-back up the complete protected data root. The requirements matrix separates preserved capabilities and limitations.