Skip to content

InvokeAI 3.0.0

Choose a tag to compare

View all tags
@lstein lstein released this 19 Jul 19:33
· 12359 commits to main since this release
v3.0.0
19cdd5a

InvokeAI Version 3.0.0

InvokeAI is a leading creative engine built to empower professionals and enthusiasts alike. Generate and create stunning visual media using the latest AI-driven technologies. InvokeAI offers an industry leading Web Interface, interactive Command Line Interface, and also serves as the foundation for multiple commercial products.

InvokeAI version 3.0.0 represents a major advance in functionality and ease compared with the last official release, 2.3.5.

Please use the 3.0.0 release discussion thread, for comments on this version, including feature requests, enhancement suggestions and other non-critical issues. Report bugs to InvokeAI Issues. For interactive support with the development team, contributors and user community, you are invited join the InvokeAI Discord Server.

To learn more about InvokeAI, please see our Documentation Pages.

What's New in v3.0.0

Quite a lot has changed, both internally and externally.

Web User Interface:

  • A ControlNet interface that gives you fine control over such things as the posture of figures in generated images by providing an image that illustrates the end result you wish to achieve.
  • A Dynamic Prompts interface that lets you generate combinations of prompt elements.
  • Preliminary support for Stable Diffusion XL the latest iteration of Stability AI's image generation models.
  • A redesigned user interface which makes it easier to access frequently-used elements, such as the random seed generator.
  • The ability to create multiple image galleries, allowing you to organize your generated images topically or chronologically.
  • An experimental Nodes Editor that lets you design and execute complex image generation operations using a point-and-click interface. To activate this, please use the settings icon at the upper right of the Web UI.
  • Macintosh users can now load models at half-precision (float16) in order to reduce the amount of RAM.
  • Advanced users can choose earlier CLIP layers during generation to produce a larger variety of images.
  • Long prompt support (>77 tokens).
  • Memory and speed improvements.

The WebUI can now be launched from the command line using either invokeai-web (preferred new way) or invokeai --web (deprecated old way).

Command Line Tool

The previous command line tool has been removed and replaced with a new developer-oriented tool invokeai-node-cli that allows you to experiment with InvokeAI nodes.

Installer

The console-based model installer, invokeai-model-install has been redesigned and now provides tabs for installing checkpoint models, diffusers models, ControlNet models, LoRAs, and Textual Inversion embeddings. You can install models stored locally on disk, or install them using their web URLs or Repo_IDs.

Internal

Internally the code base has been completely rewritten to be much easier to maintain and extend. Importantly, all image generation options are now represented as "nodes", which are small pieces of code that transform inputs into outputs and can be connected together into a graph of operations. Generation and image manipulation operations can now be easily extended by writing a new InvokeAI nodes.

Installation / Upgrading

Installing using the InvokeAI zip file installer

To install 3.0.0 please download the zip file at the bottom of the release notes (under "Assets"), unpack it, and then double-click to launch the script install.sh (Macintosh, Linux) or install.bat (Windows). Alternatively, you can open a command-line window and execute the installation script directly.

If you have an earlier version of InvokeAI installed, we strongly recommend that you install into a new directory, such as invokeai-3 instead of the previously-used invokeai directory. We provide a script that will let you migrate your old models and settings into the new directory, described below.

InvokeAI-installer-v3.0.0.zip

Upgrading in place

All users can upgrade from the 3.0 beta releases using the launcher's "upgrade" facility. If you are on a Linux or Macintosh, you may also upgrade a 2.3.2 or higher version of InvokeAI to 3.0 using this recipe, but upgrading from 2.3 will not work on Windows due to a 2.3.5 bug (see workaround below):

  1. Enter the 2.3 root directory you wish to upgrade
  2. Launch invoke.sh or invoke.bat
  3. Select the upgrade menu option [9]
  4. Select "Manually enter the tag name for the version you wish to update to" option [3]
  5. Select option [1] to upgrade to the latest version.
  6. When the upgrade is complete, the main menu will reappear. Choose "rerun the configure script to fix a broken install" option [7]

Windows users can instead follow this recipe:

  1. Enter the 2.3 root directory you wish to upgrade
  2. Launch invoke.sh or invoke.bat
  3. Select the "Developer's console" option [8]
  4. Type the following command:
pip install "invokeai @ https://github.com/invoke-ai/InvokeAI/archive/refs/tags/v3.0.0.zip" --use-pep517 --upgrade

This will produce a working 3.0 directory. You may now launch the WebUI in the usual way, by selecting option [1] from the launcher script

After you have confirmed everything is working, you may remove the following backup directories and files:

  • invokeai.init.orig
  • models.orig
  • configs/models.yaml.orig
  • embeddings
  • loras

To get back to a working 2.3 directory, rename all the '*.orig" files and directories to their original names (without the .orig), run the update script again, and select [1] "Update to the latest official release".

Migrating models and settings from a 2.3 InvokeAI root directory to a 3.0 directory

We provide a script, invokeai-migrate3, which will copy your models and settings from a 2.3-format root directory to a new 3.0 directory. To run it, execute the launcher and select option [8] "Developer's console". This will take you to a new command line interface. On the command line, type:

invokeai-migrate3 --from <path to 2.3 directory> --to <path to 3.0 directory>

Provide the old and new directory names with the --from and --to arguments respectively. This will migrate your models as well as the settings inside invokeai.init. You may provide the same --from and --to directories in order ot upgrade a 2.3 root directory in place. (The original models and configuration files will be backed up.)

Upgrading using pip

Once 3.0.0 is released, developers and power users can upgrade to the current version by activating the InvokeAI environment and then using:

pip install --use-pep517 --upgrade InvokeAI

You may specify a particular version by adding the version number to the command, as in:

pip install --use-pep517 --upgrade  InvokeAI==3.0.0

To upgrade to an xformers version if you are not currently using xformers, use:

pip install --use-pep517 --upgrade InvokeAI[xformers]

You can see which versions are available by going to The PyPI InvokeAI Project Page

Getting Started with SDXL

Stable Diffusion XL (SDXL) is the latest generation of StabilityAI's image generation models, capable of producing high quality 1024x1024 photorealistic images as well as many other visual styles. As of the current time (July 2023) SDXL had not been officially released, but a pre-release 0.9 version is widely available. InvokeAI provides support for SDXL image generation via its Nodes Editor, a user interface that allows you to create and customize complex image generation pipelines using a drag-and-drop interface. Currently SDXL generation is not directly supported in the text2image, image2image, and canvas panels, but we expect to add this feature as soon as SDXL 1.0 is officially released.

SDXL comes with two models, a "base" model that generates the initial image, and a "refiner" model that takes the initial image and improves on it in an img2img manner. For best results, the initial image is handed off from the base to the refiner before all the denoising steps are complete. It is not clear whether SDXL 1.0, when it is released, will require the refiner.

To experiment with SDXL, you'll need the "base" and "refiner" models. Currently a beta version of SDXL, version 0.9, is available from HuggingFace for research purposes. To obtain access, you will need to register with HF at https://huggingface.co/join, obtain an access token at https://huggingface.co/settings/tokens, and add the access token to your environment. To do this, run the InvokeAI launcher script, activate the InvokeAI virtual environment with option [8], and type the command huggingface-cli login. Paste in your access token from HuggingFace and hit return (the token will not be echoed to the screen). Alternatively, select launcher option [6] "Change InvokeAI startup options" and paste the HF token into the indicated field.

Now navigate to https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9 and fill out the access request form for research use. You will be granted instant access to download. Next launch the InvokeAI console-based model installer by selecting launcher option [5] or by activating the virtual environment and giving the command invokeai-model-install. In the STARTER MODELS section, select the checkboxes for stable-diffusion-xl-base-0-9 and stable-diffusion-xl-refiner-0-9. Press Apply Changes to install the models and keep the installer running, or Apply Changes and Exit to install the models and exit back to the launcher menu.

Alternatively you can install these models from the Web UI Model Manager (cube at the bottom of the left-hand panel) and navigate to Import Models. In the field labeled Location type in the repo id of the base model, which is stabilityai/stable-diffusion-xl-base-0.9. Press Add Model and wait for the model to download and install. After receiving confirmation that the model installed, repeat with stabilityai/stable-diffusion-xl-refiner-0.9.

Note that these are large models (12 GB each) so be prepared to wait a while.

To use the installed models you will need to activate the Node Editor, an advanced feature of InvokeAI. Go to the Settings (gear) icon on the upper right of the Web interface, and activate "Enable Nodes Editor". After reloading the page, an inverted "Y" will appear on the left-hand panel. This is the Node Editor.

Enter the Node Editor and click the Upload button to upload either the SDXL base-only or SDXL base+refiner pipelines (right click to save these .json files to disk). This will load and display a flow diagram showing the (many complex) steps in generating an SDXL image.

Ensure that the SDXL Model Loader (leftmost column, bottom) is set to load the SDXL base model on your system, and that the SDXL Refiner Model Loader (third column, top) is set to load the SDXL refiner model on your system. Find the nodes that contain the example prompt and style ("bluebird in a sakura tree" and "chinese classical painting") and replace them with the prompt and style of your choice. Then press the Invoke button. If all goes well, an image will eventually be generated and added to the image gallery. Unlike standard rendering, intermediate images are not (yet) displayed during rendering.

Be aware that SDXL support is an experimental feature and is not 100% stable. When designing your own SDXL pipelines, be aware that there are certain settings that will have a disproportionate effect on image quality. In particular, the latents decode VAE step must be run at fp32 precision (using a slider at the bottom of the VAE node), and that images will change dramatically as the denoising threshold used by the refiner is adjusted.

Also be aware that SDXL requires at 6-8 GB of VRAM in order to render 1024x1024 images and a minimum of 16 GB of RAM. For best performance, we recommend the following settings in invokeai.yaml:

precision: float16
max_cache_size: 12.0
max_vram_cache_size: 0.0

Known Bugs in 3.0

This is a list of known bugs in 3.0 as well as features that are planned for inclusion in later releases:

  • On Macintoshes with MPS, Stable Diffusion 2 models will not render properly. This will be corrected in the next point release.
  • Variant generation was not fully functional and did not make it into the release. It will be added in the next point release.
  • Perlin noise and symmetrical tiling were not widely used and have been removed from the feature set.
  • Face restoration is no longer needed due to the improvent in recent SD 1.x, 2.x and XL models and has been removed from the feature set.
  • High res optimization has been removed from the basic user interface as we experiment with better ways to achieve good results with nodes. However, you will find several community-contributed high-res optimization pipelines in the Community Nodes Discord channel at https://discord.com/channels/1020123559063990373/1130291608097661000 for use with the experimental Node Editor.
  • There is no easy way to import a directory of version 2.3 generated images into the 3.0 gallery while preserving metadata. We hope to provide an import script in the not so distant future.
  • The NSFW checker (blurs explicit images) is currently disabled but will be reenabled in time for the next release.

Getting Help

For support, please use this repository's GitHub Issues tracking service, or join our Discord.

Contributing

As a community-supported project, we rely on volunteers and enthusiasts for continued innovation and polish. Everything from minor documentation fixes to major feature additions are welcome. To get started as a contributor, please see How to Contribute.

What's Changed Since 2.3.5

New Contributors

Full Changelog: v2.3.5...v3.0.0rc2

Contributors

mrwho, keturn, and 46 other contributors
Assets 3
Loading