InvokeAI

What is InvokeAI?

InvokeAI is an open-source generative AI studio for creators, designed for creating images and videos. It offers functions such as text-to-image generation, image-to-image generation, local retouching, image expansion, layered canvases, ControlNet, LoRA, model management, and node workflows, all within a professional web interface; it can also run entirely on one’s own computer or server.

Invoke was initially developed by a combination of business teams and the community. The founding team joined Adobe in 2025, and the official hosting platform was shut down in October of the same year;

The maintenance of open-source projects has been entrusted to Lincoln Stein, Vic, and the community, with active development of the core code continuing.

Text-to-image and image-to-image

Users can enter positive and negative prompts, and select a model, resolution, sampler, number of steps, CFG value, and seed to generate images. In image-to-image generation, the initial image is used to determine the colors and structure of the resulting image, while the noise reduction level determines the extent to which the original features are preserved.

Different models respond differently to prompt syntax and parameters. A higher number of steps does not necessarily lead to an improvement in quality; it may even increase the generation time. It is necessary to test them according to the model’s documentation.

Advanced prompt instructions

Invoke supports prompt weights, combinations, and dynamic prompts; users can adjust the influence of various concepts or apply multiple variations at once. Prompt templates are useful for specifying the subject, style, quality, and aesthetic aspects, but it should not be done by mechanically repeating words like “high quality”.

Layered Canvas

Unified Canvas offers layer-based AI editing; users can draw, move, scale, and combine Raster Layers, and create guides, reference images, and local redraw masks for different areas. Each layer can be manipulated independently, which brings the experience closer to that of professional image editing processes.

The Generation Bounding Box in the canvas determines the context that the model can see as well as the output area; content outside this box will not be taken into account in the current generation process. When expanding the image or continuing a texture, it is necessary to include sufficient context within this box.

Local redrawing and image expansion

Inpainting focuses on regenerating only the masked area, and is suitable for correcting features in people, replacing objects, and fixing details; Outpainting, on the other hand, extends the image beyond its original boundaries.

Edge fusion depends on the mask, context, and model capabilities.

Regional hints and reference diagrams

Regional Guidance allows different descriptions to be used for various areas of the canvas, while Regional Reference Image provides local visual references. This enables users to control the position of characters, clothing, objects, and the overall composition with greater precision.

ControlNet

Invoke offers various ControlNet control parameters such as depth, edge, and pose, which are used to influence the generation process. Multiple controls can be combined, but excessive weights will result in a stiff appearance in the image, while too low weights prevent the proper maintenance of the composition.

LoRA and Textual Inversion

Model Manager allows for the installation and management of LoRA, Textual Inversion, ControlNet, and Checkpoints. LoRA is useful for adding styles, character traits, or product features, and multiple LoRA models can be combined in a weighted manner.

Each model file comes with its own license and trigger words; the Apache license for open-source software does not apply to the model weights. A separate review is required before using them in commercial contexts.

Supported base models

Invoke supports Flux, SDXL, Stable Diffusion 1.5, as well as various other open-source or open-weight models that are continuously being added. The current documentation also covers models such as Ideogram 4, Flux 2, Krea-2, and Z-Image; the availability of these models varies depending on the version and hardware used.

Some models require acceptance of the terms on the model platform as well as a configuration access token; in addition, their use for non-commercial purposes may be restricted. The status “installable” in Model Manager does not mean that the model can be used freely for commercial purposes.

Model Manager – Model management

Users can install models using the Hugging Face repository ID, file address, local path, or by scanning folders. The Manager identifies the Checkpoint and Diffusers structures, and allows for the recording of the source, description, and metadata.

Models typically occupy several GB to dozens of GB; it is necessary to plan for disk storage, backups, and handling of duplicate files. Access tokens can only be stored in the configuration settings, not within workflows or screenshots.

Node workflow

Workflow Builder uses visual nodes to create reusable pipelines; these nodes are referred to as Invocations within the Invoke function. Workflows can incorporate elements such as model loading, conditions, sampling, image processing, saving, and custom logic.

Form Builder can convert the key parameters of complex diagrams into simplified forms, making it easier for team members to use without the need to go through the node diagram to make changes.

Running workflows on Canvas

The current version allows workflows to be executed directly on the canvas’s Raster Layer. Such workflows require the Form Builder to be enabled, at least one image input, and a Canvas Output node.

The result will be placed in the canvas temporary storage area, where it can be approved by the user after review.

Custom node

Developers can use Python to create the input/output logic and execution logic for invocations, in order to develop new models, implement preprocessing, perform image operations, or integrate external services. Custom nodes run within server processes, and their source code should be checked just as carefully as that of any other code components.

Queueing and batch generation

Invoke can add generation tasks to a queue, and enable rapid iteration through batch processing, dynamic feedback, as well as fixed or random seeds. In a production environment, it is necessary to set limits on concurrency, disk cleanup, retry attempts for failures, and user resources.

Boards image management

Boards are used to organize and generate images; they allow for classification by project, stage, or type, enabling bulk downloading and management of assets. The older version of commercial hosting differentiated between private and shared Boards.

The specific multi-user capabilities of the open-source version depend on the deployment method and version.

Video generation

Invoke currently offers experimental support for Wan 2.2 video capabilities, including text-to-video and image-to-video generation, as well as models that can handle both types of tasks. The official workflow enables the creation of short MP4 clips, and Lightning LoRA is used to compress certain large models to around 4 steps.

Video models require a large amount of disk space and video memory. According to the official information, the two introductory Wan models occupy approximately 36 GB and 32 GB respectively, while installing them together – due to the shared components – requires around 56 GB of storage space.

In low-specification environments, a smaller quantization model can be chosen.

Windows, macOS, and Linux

Invoke Launcher is compatible with Windows, macOS, and Linux. For Windows, version 10 or higher is required, along with a compatible GPU.

macOS requires Apple Silicon; Intel Macs are not compatible. Linux supports NVIDIA and AMD GPUs.

Hardware requirements

The hardware requirements vary depending on the model and the output resolution. The current official guidelines suggest at least around 4GB of video memory for SD1.5, around 8GB for SDXL, and around 10GB for Flux 1.

Larger Flux, Krea, and Wan models may require 12 to 24 GB or more of video memory, as well as 32 GB of system memory.

It can run on Apple Silicon; it is recommended to have at least 16GB of unified memory. AMD GPUs are primarily supported on Linux.

The CPU mode can be started, but local generation is usually slow.

Python, Docker, and manual installation

Advanced users can install it manually from GitHub or use Docker. The current documentation requires Python 3.11 to 3.12.

Docker facilitates the isolation of dependencies and the deployment of servers, but GPU drivers and model directories still need to be mapped correctly.

Local and privacy

When using local open models, the prompts, images, and generated results remain on your own device; there is no need to upload them to third-party cloud services. When invoking API models such as GPT Image and Gemini, the data is still sent to the respective providers and is subject to their policies.

API integration with servers

The web interface for Invoke relies on local servers and nodes to handle the backend operations; developers can create automated workflows, queues, and asset management systems using these elements. When deploying it on the public internet, it is necessary to implement authentication, HTTPS, a reverse proxy, file isolation, and resource quotas.

Do not expose the default development services directly to the Internet, nor hardcode API keys in custom nodes.

Comparison of current deployment methods and costs

Invoke does not currently offer any official cloud subscription plans of its own. The prices for Starter, Indie, Premier, and Enterprise listed in the older documentation are no longer valid as Hosted has been discontinued, and therefore should not be used as current pricing references.

The fees for third-party hosting are based on the information provided on the respective providers’ official websites.

Why was the old Hosted platform shut down?

According to the official application page, after the founding team joined Adobe, the Invoke hosting service was discontinued in October 2025. The open-source version continues to be maintained by the community; users can download and install it or opt for third-party GPU hosting services.

Apache-2.0 license and GitHub

The official GitHub repository of InvokeAI makes the complete core code available; it is licensed under the Apache-2.0 license, which permits modification, distribution, and commercial use, with the requirement to retain the license and related statements.

The license applies only to the Invoke code; rights regarding the models, training data, custom nodes, and the generated content must be obtained separately.

InvokeAI usage guide

Complete a basic task.

1. In InvokeAI, specify the trigger conditions, input data, and the desired action.
2. Connect to the required application using a test account and grant only the minimum necessary permissions;
3. Configure text-to-image and image-to-image functions as well as advanced prompt parameters, mapping each field accordingly;
4. Add branch handling, failure management, retry options, and manual approval;
5. Run with a small amount of test data and check the output at each step;
6. Enable the official process only after confirming that there are no duplicate writes or abnormal charges.

Create reusable professional workflows

1. Choose business processes with a high frequency of repetition and clear rules;
2. Draw the paths for triggering, decision-making, execution, and rollback;
3. Break down text-to-image and image-to-image generation, as well as advanced prompts and layered Canvas canvases, into reusable modules;
4. Set log settings, budgets, timeout values, and permission limits;
5. Assign a responsible person to randomly check the results and handle the failed queues;
6. Review application permissions, business rules, and actual costs on a monthly basis;

Which users is it suitable for?

• Creators who wish to run AI-based image generation locally and have control over the data;
• Illustrators and designers who need layered local editing and image expansion;
• Teams that use ControlNet and LoRA to control a brand’s visual identity;
• Technical art for creating a node-based, repeatable image pipeline;
• Developers who need to set up their own generation servers;
• Open-source community members interested in researching open models and contributing code.

Product advantages

• Apache-2.0 is open-source, free, and can be hosted independently;
• Layered Canvas is more suitable for professional editing than a single-form approach.
• The node workflow and Form Builder cater to both advanced and regular users;
• Complete management of models, LoRA, and ControlNet;
• It supports Windows, Apple Silicon, and Linux;
• The project continues to be maintained by the community even after its commercial hosted phase comes to an end.

Restrictions and Precautions

• The official hosted service is no longer available; users must prepare their own local hardware or use third-party hosting services.
• Large models require significant download sizes and substantial video memory, and AMD and macOS support vary depending on the model;
• Open source does not mean that the model can be used for commercial purposes; additional security measures are required for deployment on the public internet.
• The model package for experimental video features is quite large, and the speed of generation as well as stability depend on the hardware.

Frequently Asked Questions

Is InvokeAI free?

It is free; the current core local version is licensed under Apache-2.0. Costs apply for computing hardware or third-party cloud GPUs.

Is the official InvokeAI cloud still available?

No, the original Invoke Hosted service was discontinued in October 2025; currently, the official recommendation is to use local installation or third-party hosting solutions.

Which systems does InvokeAI support?

It is compatible with Windows 10 and later, Apple Silicon macOS, and Linux; Intel Macs are not supported.

Can InvokeAI generate videos?

It offers experimental support for Wan 2.2 video generation from text and video generation from images; this requires large model packages as well as sufficient video memory.

Is InvokeAI open source?

It is open source; the core repository is licensed under the Apache-2.0 license.