Create, manage, refine, and present locally generated AI artwork.

Public Demo and Licensing

The combined Z-Image Turbo/Krea 2 public Demo provides the complete local generation workflow. Its restriction is a translucent tiled 65West logo written into every completed image. While the application is not activated, the header explains that activating with a purchased license key removes the Demo watermark from future generated images. This note disappears immediately after successful activation. The header also shows the current entitlement state and provides a silver Manage License button.

• The watermark is embedded in ZT and K2 main completed images, Session History images, preserved original-size variants, manual saves, and automatic saves.
• Images created in Demo mode are not retroactively changed after activation.
• License activation, validation, and deactivation use the production 65West licensing service and signed offline certificates.
• The numeric PIN remains a local access-control feature and is not a commercial authorization mechanism.
• The installer still displays the Terms of Use and requires acceptance before setup continues.
• Models, VAEs, LoRAs, outputs, settings, generated images, and other creative files remain locally managed.

Quick Start

Interface Overview

Header controls

Layout controls

Expand Viewer gives more horizontal space to the image area. Balanced Layout restores the standard split between options and viewer. The divider can also be dragged manually, and the program remembers the layout.

Program Functions panel order

The left panel is divided into workflow groups. Interface Style is now the first function, directly above the Prompt, and every existing control retains the same behavior and saved settings.

Each framed function uses a subtly lighter surface than the surrounding main panel. The surface color follows the selected Interface Style automatically, making related controls easier to distinguish without changing their layout.

65West Theme is the official branded presentation and remains the default. It uses the master palette from the supplied brand sheet: near-black navy surfaces, electric cyan framing, blue/violet/magenta accents, metallic-silver identity text, and the official 65West application icon. Other Interface Style choices remain available and continue to preserve the same control layout.

The Program Setup and Utility headings each include a Hide functions checkbox. Checking it collapses all functions in that group while keeping the heading and checkbox visible. Unchecking it restores the group, and both choices are remembered between launches.

Model-loading status: A wide model-optimization notice and animated activity bar appear in the center of the main interface while model, VAE, or LoRA preparation is underway. When a supported external Krea 2 BF16 checkpoint needs its one-time conversion, the notice identifies the checkpoint, its size, estimated temporary disk-space requirement, reusable Diffusers cache, and the current validation, conversion, cache-writing, or component-loading stage. The activity bar intentionally does not display an ETA or percentage because the underlying conversion and serialization operations do not report reliable completion data. During optimization, all other interface controls are temporarily locked and cannot be selected. The locked interface is covered by a neutral semi-transparent black fade instead of bright disabled-control colors. Each control returns to its previous state automatically when optimization finishes or fails.

Prompting

The prompt box appears directly below Interface Style in the left options panel. Z-Image Turbo responds best when the prompt describes what should be visible rather than relying on a traditional negative prompt.

A small Clear Prompt button sits directly beneath the prompt input and instantly empties the prompt text box. Right-click inside the prompt editor for Cut, Copy, Paste, and Select All. Keyboard users can open the same menu with Shift+F10.

Prompt Profiles

The Prompt Profile controls appear at the top of the Prompt group. They provide reusable snapshots of the prompt and the settings that affect generation.

Profiles are kept as versioned JSON files in the application’s prompt_profiles folder. Loading a profile validates referenced local resources. A missing LoRA, upscaler, wildcard folder, or face folder is identified in a warning; the unavailable item is skipped while compatible settings continue to load. Random Faces is disabled automatically when its saved folder is unavailable.

Example

A single adult female explorer, full-body standing pose, practical fitted expedition clothing, leather boots, holding an antique map, ancient jungle temple entrance, centered composition, warm rim lighting, detailed cinematic environment

Prompt emphasis

Automatic1111-style prompt emphasis is enabled by default. The Enable Automatic1111 prompt weighting checkbox directly below the Prompt editor controls the feature, and the selected setting is remembered. Clear the checkbox when parentheses or brackets should be treated as ordinary prompt characters. When disabled, prompts use their literal text without emphasis parsing.

When the checkbox is enabled, the help line below the Prompt editor summarizes the supported syntax. Emphasis is applied after wildcard, inline-choice, face-reference, and LoRA-tag processing using the active model family’s dedicated text-encoding path.

Effective weights are limited to the safe range of 0.0× through 3.0×. Excessive weights can reduce image quality, so values around 1.1× through 1.5× are the recommended starting range. If the active tokenizer cannot map an emphasized phrase safely, generation continues with the emphasis punctuation removed and records the fallback in the generation metadata instead of stopping the program.

Clear the checkbox to use literal parentheses and brackets in a prompt without emphasis parsing.

Inline choices

Use braces to randomly select one option during prompt expansion. Nested choices are supported.

A {red|blue|green} jacket, {sunrise|overcast afternoon|night lighting}

Wildcards

Place text files in the selected wildcard folder and call them with double underscores:

portrait of __character_type__ in __location__

Face-reference tag

Automatic face selection can be requested directly in the prompt:

<face_ref:Sarah.png>
<face_ref:Sarah>

The program searches the configured Prompt face-ref folder recursively and removes the tag before sending the cleaned prompt to the model.

LoRA tags

<lora:style_name:0.8>
<lora:style_name:{0|0.5|1.0}>

Models, VAE & Memory

Model directory and Active Model

The Model directory provides a local library that can be scanned with Refresh. If the field is blank, the program automatically uses the Models folder in the application’s root directory. Use Browse beside the Model directory field to select a different model library folder. Choose the model to load from the full-width Active Model pulldown, which can also accept a model path or supported repository identifier.

Keep the active external model and both protected defaults in the Transformers cache is an optional disk-space setting and is on by default. When enabled, the program waits until a replacement model has loaded successfully, preserves the active standalone-model cache, and permanently preserves caches owned by both setup-installed defaults: Tongyi-MAI/Z-Image-Turbo and krea/Krea-2-Turbo. Other inactive app-generated cache-key folders may be removed. The installed model folders under Models, original external checkpoints, VAEs, and LoRAs are never removed. Full memory reset releases only runtime objects from GPU and system memory.

Full memory reset before loading a different model remains a visible preference. In the current worker-based runtime, a genuinely different Active Model starts a fresh worker process regardless of this checkbox, which releases the old process’s system and GPU allocations. Switching models may take longer because compatible components are loaded again. No installed model, VAE, image, prompt, setting, or disk cache is deleted.

A freshly selected model can take time to preload in its worker. Loading failures are reported in the interface and in logs.

VAE

Fresh setup installs and verifies separate official VAEs for both model families: VAE\65West_Default_ZImageTurbo_VAE and VAE\65West_Default_Krea2_VAE. Changing the Active Model between Z-Image Turbo and Krea 2 automatically changes the visible Active VAE to the matching managed default before the model is loaded. Choosing a custom VAE directory refreshes the selector and clears Active VAE without automatically loading a replacement. A compatible custom VAE may still be selected manually. If the matching managed default is unavailable, blank Active VAE safely falls back to the VAE included with the selected pipeline. Run Install_Default_VAE.bat to repair both default VAEs without reinstalling the rest of the application.

Memory modes

LoRAs & Wildcards

Included LoRAs panel

Use Browse/Add to add one or more LoRAs. If the LoRA folder field is blank, the program automatically uses the LoRAs folder in the application’s root directory for browsing, discovery, and prompt tags. The Included LoRA list expands vertically as items are added so every listed LoRA remains visible without a vertical scrollbar. The list shows status, filename, and strength. Double-click the Status or LoRA area to toggle an item between Enabled and Disabled. Double-click its Strength cell to edit the value directly, then press Enter or click elsewhere to apply it. The accepted range is -2.0 through 2.0, and the edited value is saved and used for generation immediately.

• Hide All LoRA Options collapses every other control in the Included LoRAs group. The checkbox remains visible so the complete group can be restored. It is unchecked by default and does not change any LoRA selection, strength, or enable/disable state.
• Hide included LoRA list collapses the LoRA list and its selection controls to save vertical space. Uncheck it to restore the complete list. The selected LoRAs and their enabled/disabled states are preserved while hidden, and the preference is remembered between launches.
• Disable Included LoRAs temporarily — prompt tags remain active preserves the Included list but prevents those entries from being used.
• Ignore Prompt LoRA Tags temporarily leaves tags in the prompt text but omits their adapters during generation.
• Disable All LoRAs temporarily disables Included and prompt-tagged LoRAs without deleting either source.
• If the same file appears in both sources, the prompt-tag strength takes precedence and the model is loaded only once.
• All strengths must be finite values from -5.0 through 5.0. Strong multi-LoRA stacks are reported in the status, generation log, and PNG metadata but are not silently normalized.
• Default LoRA strength when browsing controls the initial strength assigned to newly added files.
• The former Selected added LoRA strength box and Apply Strength button are no longer needed; each row owns its editable strength value.
• Prompt-tag LoRAs continue to work from the configured LoRA folder.

Wildcard folder

Each wildcard file is a plain .txt file with one option per line. A tag such as __lighting__ reads from lighting.txt. Nested wildcards are chosen randomly; circular references and recursion errors report the wildcard chain and file involved in the generation error and log.

Generation Settings

Image Refinement (Second Pass)

Image Refinement performs an optional image-to-image second pass using the selected model family and its active pipeline components. It shares the loaded model, VAE, text encoder, scheduler, and compatible LoRAs rather than loading a duplicate base model.

The second pass uses a dedicated refinement-only prompt. Landscape and object scenes receive a neutral locked-scene instruction without human-specific identity vocabulary. Scenes that request people receive a separate identity and figure-count lock. This prevents the refinement prompt itself from encouraging new people or other narrative content.

Controls

Analog Photo profile

Analog Photo gives the refined result the character of a professionally scanned 35mm color-negative photograph while preserving the original subject and composition. It combines conservative image-to-image guidance with a restrained finishing stage at the selected output size.

This refinement profile is independent of the existing Analog Photo option in Image Style Enhancer. The enhancer influences the initial generation; the refinement profile polishes an already-generated image. They can be used separately or together.

• Adds deterministic fine emulsion-style grain rather than coarse digital noise.
• Softens highlight roll-off and applies subtle warm halation only around existing bright areas.
• Gently warms film-stock color, keeps skin and materials natural, and preserves readable shadow detail.
• Reduces excessive digital microcontrast and edge crispness without broadly blurring the image.
• Does not intentionally add scratches, dust, borders, light leaks, heavy fading, or an aged-photo effect.
• Uses the refinement seed for repeatable finishing, so cached-latent comparisons do not receive random grain differences.

Preservation and fallback behavior

• The refinement prompt explicitly locks identity, face, expression, wardrobe, pose, hands, camera, crop, background, props, object placement, text, and composition.
• The same expanded prompt, active LoRAs, seed information, sampler, scheduler, and loaded model components carry into the second pass.
• Face Swap and Face Detailer run after refinement so those tools operate on the image that will be delivered.
• If the installed Diffusers version lacks Z-Image image-to-image support, conversion fails, or the refinement pass cannot complete, the program keeps the original generation and reports the reason.
• After every completed, failed, or cancelled refinement pass, the program releases refinement-only images, callback data, tensors, and unused memory allocations while keeping the loaded model, VAE, LoRAs, and shared refinement pipeline cached.
• Before refinement begins, the program detaches the completed base image; releases first-pass latents, callback arguments, result containers, and retry objects; synchronizes CUDA; clears unused CUDA cache; and then rechecks GPU VRAM, system RAM, and Windows commit-memory headroom.
• If 1.50× does not have safe GPU headroom, the program tries 1.25× automatically. If 1.25× or the system-memory checks remain unsafe, refinement is skipped and the original image is kept.

Image Post Processing

The former Image Color Processing section is now named Image Post Processing. Its grayscale and saturation controls are unchanged, and it includes Automatic Image Enhancement, optional film grain, a user-configurable model-based final-output upscaler, and an optional final border.

• 1–35%: subtle correction.
• 36–65%: balanced general-purpose enhancement.
• 66–100%: stronger correction while retaining adaptive safeguards.

Film Grain strength

• 1–12%: subtle photographic texture.
• 13–25%: noticeable analog-film character.
• 26–50%: strong vintage or gritty treatment.
• Above 50%: intentionally heavy stylization.

Face Swap, Random Faces & Face Detailer

Prompt Face Swap

Preserve Reference Face Details is an optional control in Prompt Face Swap. It starts off, with strength set to 50%. When a reference face is used, it aligns fine skin detail from that reference and blends it into the swapped face after Automatic Face Detailer. It transfers local brightness detail rather than source lighting or skin color, and excludes eyes and lips. Small faces, strong angles, and poor alignment are skipped. This experimental option may help retain freckles and subtle markings but cannot guarantee an exact copy. Start at 50%, lower the strength if texture looks too strong, and compare with the option off using the same seed. The controls are saved in application settings and Prompt Profiles. No additional model downloads are required.

Place source-face images in the configured Prompt face-ref folder and include <face_ref:IMAGE NAME> in the prompt. The program resolves the requested image, removes the control tag before generation, selects the largest detected face in the completed image, and applies the requested identity. The former manual Face Swap checkbox and individual source-image selector are hidden because prompt tags provide the same selection directly inside each prompt.

A lightweight second-stage harmonization pass then matches local color and illumination, feathers the outer face boundary, and restores restrained skin texture so the swapped face blends more naturally into the generated image.

• The selected source face is cached, reducing repeated source-analysis work.
• The harmonization pass reuses the existing target-face detection and does not run INSwapper or diffusion a second time.
• Identity-defining eye, nose, and mouth regions are protected from original-face texture transfer.
• If harmonization cannot complete, the successful first-pass face swap is retained automatically.
• A prompt <face_ref:...> tag is the interface-accessible method for selecting and triggering Face Swap.
• Use the exact filename or the filename without its extension.
• The program may use GPU-accelerated ONNX Runtime when a CUDA provider is available, with CPU fallback.

Random Faces

The Random Faces group appears directly below Automatic Face Detailer. Select a folder containing face images, then enable Use a random face image for Face Swap during generation. The option is disabled by default.

• Each generation selects a new source image at random, including each Auto-Generate cycle.
• The thumbnail on the right updates before generation to show the exact randomly selected source image that will be used for the face swap. The filename and dimensions are intentionally omitted from the preview.
• PNG, JPG, JPEG, WEBP, and BMP files in the selected folder and its subfolders are included.
• An explicit prompt <face_ref:...> tag takes priority over Random Faces for that image.
• If Random Faces cannot select or apply a usable source image, the completed base image is retained and the failure is recorded in the application log.

Automatic Face Detailer

GFPGAN restoration runs after generation and after face swap. The fidelity weight ranges from 0 to 1:

• Lower values: stronger restoration.
• Higher values: stronger preservation of the generated identity.
• Prompt face thumbnail: When the prompt contains a valid <face_ref:...> tag, the compact viewer on the right side of this section displays the resolved reference image, filename, and dimensions. It refreshes as the prompt or Prompt face-ref folder changes and is also updated from the exact path accepted by generation and saved in Generation History metadata.

Automatic No-Swap Face Harmonization

Automatically enhance faces when no face swap is used is enabled by default. When a generation does not request Face Swap and the full Automatic Face Detailer is off, the program detects usable faces and applies a conservative GFPGAN cleanup with high identity fidelity, original skin-detail preservation, mild face-only sharpening, and a feathered 58% blend with the untouched generated face.

• It does not run when Face Swap or the full Automatic Face Detailer runs, preventing duplicate restoration.
• It is skipped when no usable face is detected.
• If the restoration model is unavailable or processing fails, the original generated image is retained.
• Automatic Image Enhancement can remain enabled because it runs afterward and does not perform facial reconstruction.

Cached Image Comparison

The separate Cached Image Comparison section appears at the end of Core Generation Settings, directly below Sampler / Scheduler. After generating a reference image, enable Reuse cached latent + last image state to test a different LoRA selection or strength against the same reference generation.

• Reuses the exact cached initial latent from the immediately preceding completed image.
• Reuses the previous expanded prompt, seed, model, VAE, dimensions, steps, sampler, scheduler, shift, and core processing settings.
• Uses the currently selected LoRAs and current strengths rather than restoring the reference image’s prior GUI LoRA stack.
• Does not reroll wildcards or inline prompt choices.
• The cached latent is kept only in memory during the current program session.

Main Viewer & Generation History

Main viewer

The completed image is scaled to fit the available viewer by default. Enable the actual-size/percentage option to inspect the image at a specified percentage and use scrollbars when needed. Double-clicking the viewer opens a separate frameless actual-size viewer in the center of the screen. Horizontal and vertical scrollbars remain hidden unless the full-resolution image exceeds the available viewer area on that axis. Hold the left mouse button on the image or custom status bar to move the viewer during the current viewing session, and use Close Viewer or Escape to close it.

Generation History

Generation History is stored in the program’s session_history folder. When closing, you can preserve it for the next launch or clear it.

Generated filenames include the Image Refinement profile and actual seed; Image Style Enhancer references are not included. A preserved base image uses the selected refinement profile plus Original. Images generated without a successfully applied refinement use No_Refinement.

Selected thumbnails keep a black background in the 65West Theme. A green SAVED badge identifies images that have already been saved to the Output Directory, including images saved automatically.

Live Generation Preview

While an image is being sampled, an optional temporary preview appears over the application. A cyan ✥ MOVE badge identifies the correct movable preview. Hold the left mouse button on the badge, preview header, or preview image and drag it anywhere across the full application window, including over the prompt, settings, history, or main-viewer areas. The preview remains inside the application bounds and remembers that location across generations and program launches until it is moved again.

• Enable Live Preview turns the feature on or off. When disabled, no preview decoding is performed and the speed selector is unavailable.
• Show prompt text in Live Preview hides or restores the prompt area without disabling the image preview. It is enabled by default and its setting is remembered.
• Current preview size: approximately 380 × 284 pixels.
• Slow updates about every 2 seconds.
• Normal updates about every 1 second.
• Fast updates about every 0.5 second.
• Real-Time attempts to update after every sampling step.
• The step count appears above the preview image.
• At completion, the preview displays the finished image at 100% progress and remains visible for about 2 seconds before the main viewer is updated.
• The overlay then disappears automatically; it also closes immediately when generation is stopped or fails.

Slideshow Viewer

The fullscreen slideshow can use either Generation History or the selected Output Directory.

• Set the number of seconds per image.
• Enable random order when desired.
• Pause, resume, randomize, and close from the fullscreen viewer controls.
• When using Generation History, the running slideshow refreshes in the background and detects newly generated images.
• Small images may be enlarged for display, with aspect ratio preserved and a moderate enlargement limit.
• The original image files are never modified by slideshow scaling.

Notes Box

The former Save As, Open Saved Image, and diagnostic-log button area has been replaced by a full-width multiline Notes box.

• Use it for prompt ideas, LoRA testing results, model observations, or general project notes.
• The notes remain available while the application is open.
• When the program closes, the notes are written to settings.json and restored at the next launch.

Files & Folders

The program creates its optional content directories when needed.

Recommended Portrait Baseline

Use this as a clean diagnostic baseline, then enable one optional feature at a time.

Troubleshooting

Applying Updates

1. Close the application.
2. Extract the supplied update ZIP.
3. Copy the included files into the installed program directory.
4. Allow Windows to replace the older files.
5. Launch the program and confirm that the updated behavior works. Maintenance updates can retain the same program version number.