Local-first product · AI image workflow

A local image studio built to stay maintainable.

Z-Image-Turbo-Next wraps a local image model in a focused creative workspace with six generation modes, persistent story batches, explicit resource control, and release-ready rollback points.

The interface image is a read-only capture of an empty workspace. It contains no work history, private paths, account details, or source repository link.

Empty Z-Image-Turbo-Next creation workspace with no private project data
LOCAL CREATIVE WORKSPACE Create · retouch · control · batch · review
24hdevelopment time
6generation modes
55 + 33Python tests + subtests
1.2.1local stable release

01 · Product system

The model became a product when the workflow gained boundaries.

The front end stays on one local origin while typed workflows coordinate text-to-image, image variation, Canny, Depth, Pose, and Inpaint. A single resource manager prevents local language assistance and image generation from competing for heavy resources.

01

Focused workspace

Creation, retouching, structural control, stories, the library, and system state remain separate but reachable.

02

Typed workflows

Each generation mode has an explicit versioned contract, validated dimensions, and visible settings.

03

Persistent queue

Story shots, attempts, output hashes, pauses, retries, and backups survive restarts in a local SQLite queue.

04

Rollback-ready release

Pinned runtimes, health checks, stop/start validation, and version tags keep the daily entry point recoverable.

02 · Verified evidence

Release claims come from versioned tests—not a polished screenshot.

Version 1.2.1 passed its full local release gate. Performance figures below are warm-run medians from one fixed local Windows GPU environment, with one cold run and three consecutive warm runs at each supported size.

18 / 18frontend unit tests
15 / 15non-live browser tests
15 / 15Gate H release checks
45.627s1920 × 1080 text-to-image median
50.915s1920 × 1080 Canny median
1536²daily-stable square preset

Generated story evidence

Fifteen planned scenes became a complete picture book.

“The Promise at the Twilight Platform” demonstrates the end-to-end story workflow: text-only planning, per-shot visible characters, editable prompts, persistent batches, local generation, and final layout. It also exposes consistency drift instead of hiding it.

Picture-book cover with an orange cat at a twilight station
Cover · stable visual premise
Picture-book page where an unintended second cat reveals consistency drift
Scene 07 · visible model drift
Picture-book reunion scene between the orange cat and its owner
Scene 14 · narrative payoff

Selected pages from a user-provided project output. No real person, private photo, account, local path, or original database is published.

03 · Publication boundary

Private source. Local models. Public proof.

The complete application source is stored in a private repository. Model weights, the portable runtime, the ComfyUI checkout, databases, inputs, outputs, logs, caches, and account state are excluded from version control and from this public case.

Private

Source repository access, local paths, formal data, work history, raw databases, account state, and device identifiers.

Local-only

Model weights, image generation, Qwen assistance, the persistent queue, and all source or output media.

Public

Product decisions, sanitized interface evidence, selected generated pages, aggregate tests, benchmark context, and known limits.

04 · Honest quality boundary

The workflow is reliable. The generative model is still probabilistic.

The application can preserve plans, prompts, queues, outputs, and rollback points. It cannot promise identical characters, scenes, composition, or style across independently sampled images.

Cross-image consistency

Character markings, scale, wardrobe, location details, and art direction can drift between story frames.

Editing precision

Turbo image-to-image is useful for guided variation, not guaranteed pixel-precise natural-language editing.

Performance context

Timing is machine- and workload-specific. Extreme 3312² output is evidence only, not a daily preset or product promise.

Delivery result

A finished local tool, with its limits still visible.

In roughly 24 hours of development, the project moved from a working local model to a maintainable creative product with explicit privacy, evidence, and rollback boundaries.

Development time was provided by the project owner; test counts and performance figures come from versioned local release evidence.

Back to top ↑