Downloads

v0.2.7 changelog

This update mainly focuses on bug fixes.

Fixes

  • fixeds tinyencoder url in the docker initial model download phase
  • fixes Unable to install Cuda backend on Ubuntu 22.04 #33
  • fixes Local img2img generation: initial images not visible in the job #36
  • fixes some settings not saved in the Settings window
  • fixes docker containers

New Contributors

  • @schorsch1976 made their first contribution in https://github.com/fszontagh/sd.cpp.gui.wx/pull/35

This pre release is compatible with the v0.2.5 & v0.2.6 versions.

The sd.cpp version changed (master-d9b5942)

The sd.cpp fbd42b6 tag fixes the old LoRA model unload issue #410

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/v0.2.6...v0.2.7

v0.2.6 changelog

Release Notes for v0.2.5 Hotfix

This hotfix addresses a critical issue in v0.2.5.

Fixes

  • Prebuilt binaries were compiled with incorrect CPU feature flags, causing SIGILL (Illegal Instruction) errors on unsupported processors

Original change log:

This update primarily focuses on the new server component of the GUI.
For more details, please see the related wiki page.

Fixes

  • Fixed an issue where the inpainting brush extended beyond the edges.
  • Simplified some CMake configurations for build and configuration time.
  • Added compiler optimizations for Windows.
  • Fixed a bug where the settings window occasionally closed incorrectly.
  • Corrected the format of some image metadata.

What's Changed

  • Closing the settings window no longer causes the main window to restart automatically. (Changing the language still requires a full application restart.)
  • Added a server selector under the "Generate" button. This is only visible if one or more servers are added in the settings.
  • The GUI layout has been redesigned.
  • Added a queue item version check. If a stored queue item’s version differs from the current GUI version, it will be ignored.
  • Renamed the .deb packages, removing the "~" character from the names.
  • The .deb package dependencies have been relaxed for better compatibility.

Prompts

  • Added LoRA and embedding highlighting. If a LoRA tag is present in the prompt but unavailable on the selected server or locally, it will be marked in red.
  • Added autocomplete for LoRAs and embeddings. (Embeddings are not checked for availability on the selected server or locally.)
  • Implemented a shortcut to adjust word weight like in Stable Diffusion WebUI:
    • Use Ctrl + Up / Ctrl + Down to increase or decrease the weight of the selected text or the word under the cursor.
  • Added a "Restore Last Prompt" button.
  • Added a "Clear Prompt" button to empty the input fields.
  • Added a "Normalize Prompt" button to remove duplicates and trim the string.

Model Manager

  • Local models are now separated from the server model list into a new parent category.
  • When selecting a remote model, the server list will automatically select the corresponding model’s server.

New Features

  • Introduced the TCP server component, which functions as a standalone part of the project.
  • Added a TCP client to the GUI to connect to the TCP server.

This version is incompatible with old queue job files, but outdated queue items will be automatically hidden due to the version check.

If you encounter errors on startup, please delete all configurations for a clean start.

The sd.cpp version changed (master-d9b5942)

Full Changelog: Compare v0.2.4...v0.2.6

v0.2.5 changelog

This update primarily focuses on the new server component of the GUI.
For more details, please see the related wiki page.

Fixes

  • Fixed an issue where the inpainting brush extended beyond the edges.
  • Simplified some CMake configurations for build and configuration time.
  • Added compiler optimizations for Windows.
  • Fixed a bug where the settings window occasionally closed incorrectly.
  • Corrected the format of some image metadata.

What's Changed

  • Closing the settings window no longer causes the main window to restart automatically. (Changing the language still requires a full application restart.)
  • Added a server selector under the "Generate" button. This is only visible if one or more servers are added in the settings.
  • The GUI layout has been redesigned.
  • Added a queue item version check. If a stored queue item’s version differs from the current GUI version, it will be ignored.
  • Renamed the .deb packages, removing the "~" character from the names.
  • The .deb package dependencies have been relaxed for better compatibility.

Prompts

  • Added LoRA and embedding highlighting. If a LoRA tag is present in the prompt but unavailable on the selected server or locally, it will be marked in red.
  • Added autocomplete for LoRAs and embeddings. (Embeddings are not checked for availability on the selected server or locally.)
  • Implemented a shortcut to adjust word weight like in Stable Diffusion WebUI:
    • Use Ctrl + Up / Ctrl + Down to increase or decrease the weight of the selected text or the word under the cursor.
  • Added a "Restore Last Prompt" button.
  • Added a "Clear Prompt" button to empty the input fields.
  • Added a "Normalize Prompt" button to remove duplicates and trim the string.

Model Manager

  • Local models are now separated from the server model list into a new parent category.
  • When selecting a remote model, the server list will automatically select the corresponding model’s server.

New Features

  • Introduced the TCP server component, which functions as a standalone part of the project.
  • Added a TCP client to the GUI to connect to the TCP server.

This version is incompatible with old queue job files, but outdated queue items will be automatically hidden due to the version check.

If you encounter errors on startup, please delete all configurations for a clean start.

The sd.cpp version changed (master-d9b5942)

Full Changelog: Compare v0.2.4...v0.2.5

v0.2.4 changelog

Fixes

  • In model manager, when a model set/removed as [F]avorite, the selection jumped to a wrong model
  • Improved the image management on queue-d jobs, now images separated by types to better identification
  • When an image selected eg. a txt2img to send into img2img, after the job start the original image deleted
  • fixed some GUI lagging too
  • cpu_features component now can be used from the installed system packages if exists
  • fixed the job status text on the widget
  • fixed exif / png metadata reading, now works on all supported platforms
  • output image metadatas (from any supported OS) now compatible with CivitAi
  • in the output filename format the month was one character, now this is padded with 0

Changes

  • the GUI layout is changed, trying to fix it on smaller desktop resolutions or higher / lower DPI
  • in the queue job list, now possible to select multiple item and delete them
  • in the queue job list, if you select one or more item you can delete this items with the "delete" key
  • removed ROCM build, added VULKAN. On systems with AMD GPU, the GUI use the VULKAN backend
  • Added some metadata to the binaries on Windows for better security
  • Updated localization files.

New Features

  • inpaint / outpaint and built-in mask editor (see wiki page)

The sd.cpp version changed (master-dcf91f9)

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/v0.2.3...v0.2.4

v0.2.3 changelog

Fixes

  • Resolved an issue in the new config system where certain parameters were not reloaded into the settings window.
  • Fixed an issue where the application's temporary folder was not created on the first launch.
  • Included missing shared libraries in the AppImage.
  • Addressed graphical lags in scrolled panels when using the GTK backend.
  • Fixed certain context menus for images.
  • Resolved an issue where model metadata was lost after changing a model’s path. (If a model is moved to a new path, it must be reloaded in the GUI.)

Changes

  • Replaced the old model manager’s list view with wxTreeListCtrl.
  • Redesigned the main GUI layout.
  • Simplified validation by introducing a single "Job Start" button.
  • Non-checkpoint models are now treated as checkpoint models, enabling hashing and CivitaI metadata checks for these models.
  • Removed model search and filter features.
  • Updated localization files.

New Features

Model Manager

  • Introduced the "Model Favorite" feature:
    • Allows users to mark models as favorites. You can configure the main GUI to display only favorite models—useful for managing a large number of models.
  • Models are now categorized by type (e.g., Checkpoint, LoRA, ControlNet, etc.).
  • Subfolders in the model manager now act as categories. For example, creating a folder named SD1x and moving models into it will separate those models (e.g., CHECKPOINTS/SD1x).
  • Added the ability to delete model files directly from the GUI, which also removes their metadata.

Generation Settings

  • Added prompt presets (styles/templates). See the GitHub Wiki for more details.

Build system

  • New cmake parameters:
    • WXWIDGETS_EXTRA_PATH: additional path, where wxWidgets installation will be searched #24
    • WXWIDGETS_VERSION: this forces cmake to search for the specified version of the wxWidgets. Defaults: 3.2.6. #24

The sd.cpp version not changed (master-b5f4932)

Full Changelog: v0.2.2...v0.2.3

v0.2.2 changelog

This update addresses issues present in version 0.2.1:

  • Fixed a deadlock issue that caused the GUI to freeze, preventing queue jobs from completing once started.
  • Resolved the missing model hash information in the metadata of output images.
  • Corrected some metadata reading issues for images in the Image Info tab

UPDATE:

  • Added a desktop widget, now available for both Windows and Linux.

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/v0.2.1...v0.2.2

v0.2.1 changelog

This update primarily includes minor, less visible fixes and some new GUI features.

Changes

  • The secret store name has changed, requiring you to re-enter your CivitAi API key.
  • Fixed CMake dependency checks to make it easier to build from source on a wider variety of Linux distributions (see: #22 ).
  • Several small improvements were made to the CivitAi downloader, including more robust handling of HTTP errors.
  • Resolved an issue where the model name used in img2img did not appear in the queue list.
  • Queue jobs are now sorted by default so that the newest jobs appear at the top.
  • Image-related processes have been optimized.
  • The language selector in settings has been fixed to ensure each language appears only once.
  • The language detector on first startup is fixed (again)
  • Added the -vulkan parameter to the GUI (vulkan is coming)
  • Minor logging issues have been fixed.
  • Improved external process detection on both Windows and Linux.
  • Language files have been updated again.
  • Refactored config handler
  • Fixed some bug on the selectors handling (sample method, scheduler, wtype...)

New Features

  • PNG compression level can now be configured (this does not affect image quality).
  • The existing quality setting for JPG images remains the same.
  • Output image filename formatting is now customizable (e.g.: [mode]/[year]-[month]-[day]/[jobid][seed][width]x[height]). Please see the new wiki page for details.
  • Image metadata handling has been improved:
    • Generated PNG images now embed PNG metadata compatible with Stable Diffusion WebUI (PNG info).
    • EXIF data is now embedded into JPEG formats
  • Added an Image Info Tab into the GUI, where images can be checked for generation data (similar to Stable Diffusion Webui PNG Info tab). Compatible with the most image generators.
  • Added an progressbar and a status label to the main gui
  • Added an experimental floating Widget which is shows the current running job's details. This curretly only available on GTK and in non Release builds.

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/v0.2.0...v0.2.1

v0.2.0 changelog

New Features

  • New Options on the GUI:
    The following options related to keeping components on the CPU have been added to the graphical interface.
    Any changes to these options are immediately saved in the configuration file and will persist when restarting the application.

    • Keep VAE on CPU
    • Keep CLIP on CPU
    • Keep ControlNet on CPU

  • Support for New Models:
    Added support for ClipL, ClipG, and T5XXL models (currently available only on the txt2img tab). This also enables compatibility with SD3x and FLUX models. For FLUX models, the file must currently be manually selected on the txt2img interface. However, it is recommended to store these models in the same directory as other models.
    More information is available on the wiki page.

    • Queue jobs generated using these models can now also be reloaded.
    • For Diffusion models, the GUI will attempt to locate any additional required models in the VAE directory.

  • Skip Layer Guidance for SD3.5 Medium Models:
    Added support for the "skip layer guidance" feature.
    (See details: GitHub PR #451)

  • Flash Attention for Diffuser models: Added the option to use Flash Attention with the diffusion models, which may reduce the memory usage

  • Updated sd.cpp:
    Updated to version master-b5f4932.

  • FLUX 1 Lite 8B Support:
    Added support for the Flux 1 Lite 8B (Freepik) model.
    (See details: GitHub PR #474)

Several GUI updates have been made, and the language files have also been updated. This version nearly covers all the features available in the sd.cpp example program. However, minor and/or major bugs may still be present.

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/preview11...v0.2.0

preview11 changelog

This is the first preview with multi-language support. On the first start, the GUI will detect the OS locale and search for available language files. If no match is found, the fallback language is always English. To change the GUI language, you can go into the Settings -> GUI In this version, if you close the settings window, the GUI will fully reload. (All modified generation settings will be lost)

This version includes some refactoring, with a precompiled header added to the main app and a refactored main configuration logic. Due to the refactoring, you may need to delete old configuration files if the app fails to start.

Fixed some bugs:

  • After model quantization, the model list now reloads correctly
  • Fixed some GUI spacing issues when non-English words have more characters than English
  • and other QoL bugs

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/preview10v2...preview11

preview10v2 changelog

Full Changelog: https://github.com/fszontagh/sd.cpp.gui.wx/compare/preview10...preview10v2

This is a fix for the preview10:

  • fixed on windows the temp path name in the external process
  • fixed the shared library naming on windows
  • fixed the job status reading from the shared memory + changed the maximum size the shared memory to 1MB
  • fixed when a job started, the selected floating type was ignored

preview10 changelog

This release introduces several updates and improvements:

  • this release uses the latest tag from sd.cpp master-ac54e00

  • Resolution Presets: Added resolution presets to the GUI. (Note: prompt presets are not yet implemented.)

Resolution Presets

  • Model Converter: A model converter/quantization has been added to the Models list. For details, see the Wiki page.

  • CivitAi Toggle: The settings now include an option to enable or disable CivitAi features (enabled by default).

CivitAi Toggle
Disabling CivitAi features will hide related buttons and menu items, and prevent the app from downloading CivitAi information after model hash generation (no online model hash verification).

  • "Stop and Restart Background Process" Button: Added a new button for stopping and restarting the background process.

image

This feature is part of a recent refactoring. For more information, refer to the GitHub Wiki.

Major Refactoring: Background Process

The major update in this release is a background process that manages tasks and runs the queue separately from the main app, reducing the risk of segmentation faults (segfaults) in the GUI. If the background process stops unexpectedly, the main GUI will automatically restart it, enhancing overall stability.

  • Enhanced Job Statistics: Additional statistical information is now included in the job JSON files, which will be useful for future enhancements.
  • Fixed CivitAi downloader: added some checks into the model downloader. Checking if the model file already exists or the download just failed
  • Fixed character encoding: on linux and windows, if a path contained a special character (áéíőó), the string returned as empty.

There was more code enhancements in the background, but there is still much to do

Upgrading from the previous prereleases

Please before installing the new version, backup and remove the old job files from the data directory. The old job files format are incompatible with the new. Please see the wiki for the application paths.

The screenshots are created under Kubuntu 24

preview9 v2 changelog

This is the same as the preview9, only the packaging changed. Added hibplas and fixed the AppImage naming

preview9 changelog

What's Changed

  • changed sd.cpp version to newer (master-e410aeb)
  • replaced fully the build system to cmake
  • fixed nvidia detection on linux
  • fixed progressbar on civitai downloder + added check if the downloaded file is too small (eg. no permission to download a model without civitai api key)
  • added some help text + updated help urls in the ui
  • replaced installer on windows to NSIS #18
  • fixed deb dependencies
  • added AppImage version #16
  • added .tar.gz version
  • Ubuntu22, Ubuntu24 and windows support (other linux distributions may work with the AppImage / .tar.gz variations)
  • added the missing 'scheduler' selector into the gui
  • renamed the default Type from 'Default' to 'Count'. Count means: get the model type from the modelfile
  • added all available Sampler into the selector
  • added .pt extension when searching for upscaler models
  • removed the limitation to set the image target resolution
  • added more system information into the About window
  • added a disclaimer into the About window
  • added translator python script into the po directory, to generate translation with ollama

The build system change give the opporunity to use github workflow for auto builds. Currently it is working as - is.

Known failures, TODO-s:

  • need more time to improve the CivitaiDownloader part of the app (save download history, more error handling)
  • handle the assertions what is comming from sd.cpp / ggml.cpp
    • currently if not enought cuda memory or the ggml can't handle a model, the assertion occurs and the app exiting without user notification
  • currently the multilanguage feature is implemented in-half, po files are ready, but need to implement the language handler into the GUI
  • delete model files from the disk in the Models tab
  • add hipblas, vulkan versions

preview7 - 0.1.6 changelog

Added features

  • CivitAi.com model downloader/browser

    Limitations

    • download history not saved
    • when the window is closed, the download will be aborted without user notification
    • aborted downloads part file (modelname.download) will not be deleted, need to be delete manually or it will be overwritten when a new download start with the same model file
    • no auto hash calculation after downloads
    • only locally hashed models will be marked in the CivitAi downloader as locally available model (if you not hashed a model file, and you will download the same file with same name, the old model file will be overwritten without notification)

  • CivitAi api key for private model download (CivitAi docs). You can add your key in the Settings -> CivitAi tab

Modifications

  • limit target image resolutions, only allowed resolutions with dividet by 256 (except for loading settings from old images)
  • wxWidgets now statically linked into the binary

Some bug fixed in the linux versions. The deb files are created on Ubuntu 22. Now the gnome-secret / ksecretsservice etc.. is required to store the CivitAi api key. The CPU and VGA versions are now in separated .deb files.

preview7 - 0.1.5 changelog

This preview release have some gui enhancement and inner data handling refactoring. This may cause unexpected behaviours. If you experience it, please report it in a new bug on the github's issues page.

Added features

  • On windows and linux, dynamically load the library which is required / best to (currently available rocm, cuda, avx, avx2, avx512) the HW
  • fully reworked build system, no more cmake just chalet
  • on linux, the dependencies are depends from the OS, not vcpkg (supported wxwidgets 3.2 on linux)
  • preview images are now on the bottom of the lists (queue job list & model info llist)
  • some gui behaviour is just changed. When settings is open, the main ui will freezee and hide to disable unexpected behaviours on settings change
  • added the following args to override the detected HW configurations:
    • "-cuda" to force to load cuda
    • "-rocm" to force to load rocm
    • "-avx"
    • "-avx2"
    • "-avx512"
    • "-allow-multiple" to allow multiple instances to run. Eg. if you wamt to start a cuda and an avx512 instance too: sd.gui.exe -cuda -allow-multiple sd.gui.exe -avx512 -allow-multiple The first instance no need to start with the "-allow-multiple" parameter to run another new one.
  • currently the logging is set to debug in the log on the bottom. Are logs are now written in a log file into the appdata directory New sd.cpp branch appeared, but when i tested i found some bugs which is prevents to upgrade.

preview6 - 0.1.4 changelog

These preview release have some gui enhancements and inner thread signaling / data handling refactoring. This may cause unexpected behaviours. If you experience it, please report it in a new bug on the github's issues page.

Added features

  • CivitAi model info and model images previews for embeddings too image

  • Thumbnails are cached wich downloaded from CivitAi info or generated by diffusion

  • Queue jobs more details and features per image: image

  • Queue item right click features: image

  • Incompatible controls are disabled respect to the generation mode, eg. on upscaler tab the CFG is ignored

  • Added splash screen while application startup and loading

  • Only one instance allowed to run in one time image

  • Pending queue items now are possible to pause. If you set to paused, then it just ignored from the queue. Running job can not be paused (not implemented in sd.cpp)

  • Paused queue items now possible to set to pending image

  • On a queue item a Requeue clones the job, but set the Status to paused (so not started automatically) image

  • Added semi-auto versioning (see the cube button next to the Settings and Refresh button)

  • Added some place holder images to the ui

Bug fixes

  • after a txt2img job a new img2img job causing segfault
  • on model tab the model filters are works with text filter too
  • refresh models now works with the model list (on the Model tab)
  • some json converter got more validation (as always)
  • more data check on events (eg. data table index check)

Update from previous preview versions

Just run the attached installer. Stop all instances before running the installer. Queue job files (.json) partially compatible with the new format. May the old jobs images visible as "deleted". If the application won't start after upgrade, please remove all .json files

Current limitations / known bugs

  • large textual embeddings are causing crash

preview5 - 0.1.2beta changelog

These preview release have some gui design refactoring and inner thread signaling refactoring. This may cause unexpected behaviours. If you experience it, please report in a new bug on the github's issues page.

Added features

  • CivitAi model info and model image previews for Lora and Checkpoints in the model list
  • CivitAi model info and model image previews now will downloading after the model's hash check finished using own thread, so the window not freezee while it's downloading
  • App icon on Windows
  • UI redesign, resizable window
  • On job queue list now the generated and input images are visible, with double click it's opens in the OS default app
  • 10 char length sha256 (Aka Autov2 hash)
  • updated readme
  • updated the backend to the latest stabble-diffusion.cpp

Bug fixes

  • after an upscale job a new job causing segfault
  • on model tab the search is now faster, but some bug still in at filter
  • better "Refresh models" button behavior - but in model list it's make some weird things

Update from previous preview versions

Just run the attached installer. Stop all instances before running the installer. Currently the config only compatible with the preview3 and preview4 versions. Older versions config files are incompatible, need to delete before run the app. If you experience freezee on startup, please delete all .json file witch is generated by the app.

Current limitations / known bugs

  • large textual embeddings are causing crash

preview4 changelog

Added features

Upscaler limitations

  • only 4x upscaling available (sd.cpp limitation), but you can upscale an upscaled image (eg. drag and drop into the "Upscale" tab the already upscaled image :) )
  • only tested with RealESRGAN_x4plus_anime_6B.pth. The RealESRGAN_x4plus may work too (with a lot of warning log)
  • the upscaler working as tiled upscaler, so tile artifacts sometimes visible

Update from previous preview versions

Just run the attached installer. Stop all instances before running the installer. Currently the config only compatible with the preview3 version. Older versions config files are incompatible, need to delete before run the app.

Current limitations / known bugs

  • with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)

  • max 75 tokens in all prompts, more tokens are just ignored

  • when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context

  • when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context

  • CivitAi info only works on checkpoints

  • CivitAi info downloading is working on the foreground, the window freezee

  • SD.next using (10char length sha256) -- ??

preview3 changelog

Added features

Job management (Jobs and Images tab)

  • double click will show some details about the job (in an ugly window. You can resize it if you want :) )
  • now is possible to delete a job (depends on his status) Only one job at once. Just hit right click on it and select "Delete"
  • With right click on a queue item you can:
    • Requeue: this will make a clone and start the job
    • Copy to text2img: copy some parameters from the job into the current settings (no prompts will be copied)
    • Copy prompts to ... : copy the prompts from the queue job
    • Detail...: open up the same ugly window as double click. Here you can find some details about the job (nothing fancy, nothing functionality at)
  • Job items now sortable by date

GUI enhancements

  • re-added desktop notifications, configurable from the settings (turn on/off, set timeout)
  • addedd drag and drop functions:
    • text2img accept image drop-in (drag&drop). It can detect meta info from jpeg (which jpeg was generated by this gui or the SD.next)
    • img2img accept image drop-in (drag&drop). It can detect meta info from jpeg (which jpeg was generated by this gui or the SD.next)
  • The new default output image format is now jpeg with quality 95. The output images will now store the meta-info from the parameters of the image generation. This images can be drop-in into eg. img2img. It will autofill the prompts and other settings from the image. Some enhancments requied here too. I didn't tested, but in theory, this is compatible with SD.next.

Model management (Models tab)

  • Added two checkbox: Lora & Checkpoints
    • This will filter when you use the searchbox
  • Added searchbox which will filter the list by name (min. 3 char required to work). If the Lora checkbox is checked but Checkpoints not, then only Lora will be filtered. Need more enhancements :)
  • You can sort the models by name
  • Loras now in the model list too
  • right click:
    • Calculate hash: calclute the sha256 hash from the file. The hash is needed to identify the model by CivitAi. (currently the hash is not compatible with Automatic and other webuis)
    • Select model ... : this will select the model on the top list. When hit the "Queue" button somewhere, then this model will be used
  • Doubleclick:
    • On checkpoint type model, a new window will popup with some model info from CivitAi. This currently only works with Checkpoints. With lora, it wil freezee the gui and drop a segfault.
    • Currently the model info is not optimized, so the window will freezee while downloading the infos and the images. Please use with cauction

Current limitations / known bugs

  • with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)
  • max 75 tokens in all prompts, more tokens are just ignored
  • when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context
  • when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context
  • On start, the gui has no default selected quantation (this is not interfere with the normal working)
  • CivitAi info only works on checkpoints
  • CivitAi info downloading is working on the foreground, the window freezee
  • on job start if the selected model need hash, the hashing is starting, but progressbar not upgraded in the Model tab.
  • SD.next using (10char length sha256) -- ??

If you have an earlier preview release and the gui won't start, please delete all settings / presets / model info files (all is in .json files) and try again.

preview2 changelog

Current features:

  • text to image
  • image to image
  • keep model in the memory between generations (see limitations)
  • controlnet
  • optional tiled vae
  • optional TAESD encoder
  • optional VAE model
  • custom quantation, model converting on the fly
  • save / load / delete presets
  • job queue
  • progressbar
  • desktop notifications
  • lora support in the prompts (see limitations) lora:lora_filename_without_extension:weight
  • textual embedding support in prompt (see limitations)
  • auto generate sha256 from model files
  • manual sha256 generation from model files and lora files in model tab
  • clone a job
  • copy prompt from a job
  • copy settings from a job

Current limitations / known bugs

  • with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)
  • max 75 tokens in all prompts, more tokens are just ignored
  • when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context
  • when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context
  • On start, the gui has no default selected quantation (this is not interfere with the normal working)

preview1 changelog

Current features:

  • text to image
  • image to image
  • keep model in the memory between generations (see limitations)
  • controlnet
  • optional tiled vae
  • optional TAESD encoder
  • optional VAE model
  • custom quantation, model converting on the fly
  • save / load / delete presets
  • job queue
  • progressbar
  • desktop notifications
  • lora support in the prompts (see limitations) lora:lora_filename_without_extension:weight
  • textual embedding support in prompt (see limitations)

Current limitations / known bugs

  • with controlnet, the base model have to be reloaded on all new image generation (it's automatic at least)
  • max 75 tokens in all prompts, more tokens are just ignored
  • when lora is specified in the prompt from the previous job, and base model not changed, the old lora keep in context
  • when textual embedding is specified in the prompt from the previous job, and base model not changed, the previous textual embedding keep in context
  • if base model changed and lora used in previous job's prompt, the gui is just crashing