surface
kaitoi VS-1 compare
01 cmp comparison
All comparisons

Kaitoi and Invoke compared

For Invoke users who want a hosted studio in the browser, or video, a storyboard, a timeline and an assistant around their image work, without leaving their own GPU and local models behind.

Invoke, also published as InvokeAI, is a free, open-source creative engine for AI image generation under the Apache 2.0 licence. It runs on your own computer and pairs a layered canvas with a node editor, and its v7 beta adds video.

In October 2025 Adobe announced that the Invoke team had joined the company, and Invoke's hosted platform has shut down. The open-source Community Edition continues under long-time maintainers, with regular releases throughout 2026, and Invoke 7 is in public beta.

Kaitoi is a hosted creative platform in the browser, for anyone who wants that, where AI models, image tools and Python code sit in one graph beside a storyboard, a timeline and an assistant, Penny. It is free to start with no subscription, hosted models cost the provider's own price with no markup, and local work still runs through Kaitoi CLI, on your own computer or a GPU machine elsewhere.

Checked against each product's own documentation on .

02 str strengths

What Invoke does well

A layered canvas. Invoke's canvas lets you paint, compose and generate on independent layers, with inpainting and outpainting, control layers that guide composition with depth, edge or pose maps, regional prompts for different areas of an image, editable vector path layers, and export to a layered PSD.

Open source, under Apache 2.0. The code is open under a licence that permits commercial use, and Invoke says it is the foundation for several commercial products.

Fully local. Invoke runs on your own GPU, on Windows, Linux or an Apple Silicon Mac, installed with an official launcher, a Docker image or by hand for headless servers. With local models, your prompts and images stay on your machine.

Nodes and custom code. A visual node editor builds reproducible pipelines, where a workflow can call another workflow and nodes can cache their results. Developers write custom nodes in Python, the Custom Node Manager installs node packs from a Git URL inside the app, default workflows give a place to start, and a workflow can expose a simple form so others can run it without editing the graph.

Many open models, even on smaller GPUs. Invoke supports SD 1.5, SDXL, FLUX, Qwen Image and many newer open model families, and a Model Manager installs and organises checkpoints and LoRAs. A low-VRAM mode, on by default, helps large models run on smaller GPUs, and since 6.13 Invoke can also call hosted models such as GPT Image and Gemini on your own API keys.

An active community. The Community Edition has shipped regular releases throughout 2026, the docs list community-built node packs, and a community showcase and weekly art challenges share what people make.

03 apr approach

How Kaitoi approaches it

Canvas work as nodes. Inpaint, outpaint, erase, background removal, layer splitting and upscaling models run as nodes, with overlay, blend and mask nodes and a drawing canvas for sketches, masks and crops. A structure-control node from Stability AI guides a new image by the composition of an input image, Google's Nano Banana, GPT Image, Seedream, FLUX.2 and Qwen Image edit nodes change an image from a written instruction, and Recraft Vectorize, Image to SVG and QuiverAI's models make SVG. A See-through node splits an anime character image into a layered PSD. There is no layer-based editor: Kaitoi has no stacked control layers, regional prompt layers or vector layers, and SVGs are node outputs rather than an editing layer. Layered PSD export and import are coming soon.

Open source, coming soon. Kaitoi Studio is a hosted service, not open source, and Kaitoi Open Source, a distribution you run yourself, is announced as coming soon. Studio is free to start, with no subscription, and hosted models cost the provider's own price with no markup. To build on Kaitoi today, a saved template becomes an API endpoint, and a workflow becomes a Station that other people run.

Local where you want it. Kaitoi CLI runs local work on your own computer or a GPU machine elsewhere: it connects your ComfyUI, Ollama and Blender to Studio, and Runtime recipes run your own models there. Kaitoi Code, the coding agent in Kaitoi CLI, writes and tests those recipes from a Hugging Face model or a GitHub repository, and on Studio an exported ComfyUI workflow becomes one node. Steps on your own hardware use no credits, and Kaitoi never trains models on your work. Studio itself is hosted in the browser and needs a connection, and there is no desktop app.

Nodes and custom code. Every step is a node in a typed graph, also shown as a simpler Blocks chain, where Create subgraph folds a selection into one reusable node and a run reuses the results of steps that did not change. Nodes are Python, written by Penny on every plan and by hand on Studio or Enterprise, and a node can carry its own HTML, CSS and JavaScript interface, WebAssembly included. Nodes and templates publish to Community, saved templates start new projects, and a Station gives a workflow a simple form or a custom app at a stable link. Hosted nodes use Python packages from an approved list; code that needs anything else runs on your machine as a Runtime recipe.

Many models, with no GPU needed. More than 1,400 hosted generative models and around 400 language models, as of September 2026, need no GPU of your own and run on credits at the provider's own price or, for providers that accept them, on your own keys. FLUX.2, Qwen Image, Z-Image, GPT Image, Google's Nano Banana and Seedream are among the nodes, and the open checkpoints in your own ComfyUI, SD 1.5 and SDXL among them, run through Kaitoi CLI; there is no hosted SD 1.5 or SDXL node. A FLUX.2 trainer node trains LoRA weights from your images for the FLUX.2 LoRA node, and Qwen Image and other models have LoRA nodes too. The New Models panel has Penny turn a hosted model without a node into one, and Studio shows a credit estimate before a node or graph runs. Penny, a storyboard and a timeline sit beside the models in the same project.

A community of its own. Nodes and templates publish to Kaitoi's Community, where listings are reviewed, and Stations can be listed in a reviewed gallery at kaitoi.io/stations, except one that runs your own code, Penny-written nodes included, which is invite only. Read-only share links show a project to people outside your team.

Beyond images, a production. On every plan, Penny builds and edits graphs with you, writes the nodes that are missing, and edits storyboard frames, timeline clips and Lab boards, following the Skills and Playbooks you give it. LTX video models generate, extend and restyle video as nodes, beside MiniMax's Hailuo and H3 video models, Wan video models and Luma's Uni-1, Photon and Ray models, and a storyboard of scenes and shots feeds a multi-track timeline that renders MP4, WebM or MOV in ProRes and exchanges OpenTimelineIO with editors.

Reached from your own code and agents. Kaitoi's MCP server, on every plan, lets compatible AI clients build graphs, run approved work and inspect results, and on Studio write nodes. A REST API runs graphs from your own code, a saved template becomes an asynchronous API endpoint whose results are polled or streamed, and Penny can call outside MCP servers.

Working with others. On a team Studio subscription, team workspaces have roles, shared projects and pooled credits, with weekly caps per member. Edits do not sync live: when a teammate saves, Studio offers to load the newer version.

04 tbl side by side

Invoke and Kaitoi, side by side

Checked November 13, 2026.
QuestionInvokeKaitoi
Where the work runs Whose hardware runs the models, and what you install.On your own GPU, through a local web server you open in the browser, installed with the official launcher on Windows, Linux or an Apple Silicon Mac, a Docker image, or by hand. The download page links third-party GPU hosts such as RunPod and Railway.Studio is hosted and runs in the browser, on phones and touch screens too, with hosted models on credits. Kaitoi CLI connects your own computer or a GPU machine elsewhere, so your ComfyUI, Ollama, Blender and Runtime recipes run there, using no credits. There is no desktop app, and Kaitoi Open Source, a distribution you run yourself, is announced as coming soon.
Image canvas and editing Painting, masking and changing images you already have.A layered canvas with inpainting and outpainting, control layers guided by depth, edge or pose maps, regional prompts, editable vector paths and layered PSD export. The v7 beta adds layer groups, 16 blend modes and non-destructive adjustments.Inpaint, outpaint, erase, background removal, layer splitting and upscaling models run as nodes, with overlay, blend and mask nodes, a drawing canvas for sketches and masks, and colour nodes for levels, colour wheels, LUTs and white balance. A structure-control node from Stability AI guides a new image by the composition of an input image, Google's Nano Banana, GPT Image, Seedream, FLUX.2 and Qwen Image edit nodes change an image from a written instruction, and Recraft Vectorize, Image to SVG and QuiverAI's models make SVG. A See-through node splits an anime character image into a layered PSD. There is no layer-based editor, so there are no stacked control layers, regional prompt layers or vector layers. Layered PSD export and import are coming soon.
Models Which generative models you can reach, and on what hardware.SD 1.5, SDXL, FLUX, Qwen Image and many newer open model families, installed and organised in a Model Manager with checkpoints and LoRAs, and run on your GPU with a low-VRAM mode on by default. Since 6.13, GPT Image, Gemini, Seedream and Qwen models also run on your own API keys, billed by those providers.More than 1,400 hosted generative models and around 400 language models as of September 2026, needing no GPU of your own, on credits at the provider's own price or, for providers that accept them, on your own keys. FLUX.2, Qwen Image, Z-Image, GPT Image, Google's Nano Banana and Seedream are among the nodes, with LoRA nodes for FLUX.2, Qwen Image and others, and the open checkpoints in your own ComfyUI, SD 1.5 and SDXL among them, run through Kaitoi CLI. There is no hosted SD 1.5 or SDXL node. The New Models panel has Penny turn a model without a node into one.
Training Teaching a model your style or subject.Invoke Training, a separate official tool, trains LoRAs for SD and SDXL, and LoRAs apply in the generation panel, the canvas and workflows.A FLUX.2 trainer node trains LoRA weights from your images inside the graph, as a paid node through fal, for the FLUX.2 LoRA node, and Qwen Image and other models have LoRA nodes too. It is the only trainer node, so there is no SD or SDXL trainer node.
Nodes and your own code Building pipelines, and adding steps the product does not ship.A visual node editor where a workflow can call another workflow and nodes can cache their results. Custom nodes are written in Python and installed from a Git URL with the Custom Node Manager, and the docs list community node packs.A typed node graph, also shown as a simpler Blocks chain, where Create subgraph folds a selection into one reusable node and a run reuses the steps that did not change. Nodes are Python with packages from an approved list, and can carry their own HTML, CSS, JavaScript and WebAssembly interface. Penny writes nodes on every plan, editing code by hand needs Studio or Enterprise, outside agents write nodes through Kaitoi's MCP server on Studio, and nodes publish to Community. Code that needs other packages runs on your machine as a Runtime recipe.
Running work Queuing, stopping and running steps side by side.Generations run through a visible queue where jobs can be cancelled, batch generation makes many images in one go, and since 6.14 a machine with several GPUs runs queued jobs in parallel.Run one node, a selection, or every node with select all, with the upstream steps they need, and stop a run while finished steps keep their results. Independent nodes run at the same time, one output can feed several nodes, and an array input collects several results into one step; there is no node that runs a workflow once per item. Studio shows a credit estimate before a node or graph runs.
Sharing a workflow Letting other people run what you built.A workflow's simple form travels with it, so other Invoke users can run it in their own install, with the same models.Stations at a stable link: a simple form, or a custom app built from a node's own interface, that visitors run on their own Kaitoi account and pay for themselves, listed if you like in a reviewed gallery at kaitoi.io/stations. A Station that runs your own code, Penny-written nodes included, is invite only, and steps that need a local service stay in Studio. Templates start new projects and share by link.
Video and finishing Making video and putting a cut together.Experimental video: local Wan 2.2 clips through workflows since 6.14, and in the v7 beta a Video panel that makes short clips from text, images or video with Wan 2.2, MiniMax H3 or LTX-2.5, some with a soundtrack, and extends, trims and joins them.LTX video models generate, extend and restyle video as nodes, and MiniMax's Hailuo and H3 video models, and Wan video models, run as nodes too, beside Luma's Ray models and other video edit, extend, reframe and upscale nodes; ElevenLabs speech, sound effects and music are nodes as well. A storyboard feeds a multi-track timeline that renders MP4, WebM or MOV in ProRes and exports and imports OpenTimelineIO. Video workflows in your own ComfyUI join through Kaitoi CLI.
Teamwork Working with other people on the same work.An opt-in multi-user mode, since 6.12, lets several people share one Invoke server, with administrator and regular user roles, and boards and workflows that can be private, shared or public.Team workspaces with roles, shared projects and pooled credits on a team Studio subscription, weekly caps per member, and read-only share links with optional cloning. Edits do not sync live: when a teammate saves, Studio offers to load the newer version.
API and outside agents Driving the product from your own code or an AI client.Invoke's local server exposes an HTTP API, and clients run work by submitting a graph to its queue.A REST API runs graphs from your own code, and a saved template becomes an asynchronous API endpoint whose results are polled or streamed. Kaitoi's MCP server, on every plan, lets compatible AI clients build graphs, run approved work and inspect results, and on Studio write nodes, and Penny can call outside MCP servers.

Each cell comes from the product's own website, documentation or code repository. "Not publicly documented" means we found no official page that says so, not that the feature is missing. Plans and features change; check the official pages before you decide.

05 fit which to choose

Which one fits

Choose Invoke if
  • You want the whole app on your own machine, with your prompts and images staying there.
  • You want open-source software under Apache 2.0 that you can change and build products on.
  • You paint and compose on layers, with control layers, regional prompts and layered PSD export.
  • You want one app that installs and organises your local checkpoints and LoRAs, with a low-VRAM mode for smaller GPUs.
Choose Kaitoi if
  • You want a hosted studio in the browser that needs no GPU, with your own machines joining when you want them.
  • You want more than 1,400 hosted models with no subscription, at the provider's own price.
  • You want an assistant that builds graphs and writes the missing node in Python.
  • You want images, video, a storyboard and a timeline in one project, with OpenTimelineIO for your editor.
  • You want to publish a workflow as a form or an app where each visitor pays for their own runs.
06 dem example

A real example: a local model becomes a node

Kaitoi Studio film

Kaitoi Signals #4 - Local Runtime Recipes

Kaitoi CLI turns an external Hugging Face repository into a runnable local recipe, exposes its parameters inside Kaitoi, and generates a runtime node that joins the graph like any other capability.

This film, from our Kaitoi Signals series, takes a model from an outside repository and runs it on your own machine, inside the graph.

  1. Kaitoi CLI reads an external Hugging Face repository and turns it into a runnable local recipe.
  2. The recipe runs on the local machine, and its parameters appear inside Kaitoi.
  3. A runtime node is generated and joins the graph like any other node.

Read more in the documentation:

07 prc pricing

Pricing and access

Invoke

Invoke is free to download and use under its open licence. You pay for your own hardware or a third-party GPU host such as RunPod or Railway, and external API models bill your own provider account.

Kaitoi

Free to start, with no subscription. Credits pay for hosted models, including Penny's, at the provider's own price with no markup, and for paid Community nodes; steps that call neither, and your own hardware, cost nothing. Your own provider keys work on every plan and are billed by your provider. Studio adds included credits, team seats and hand-edited code.

08 faq questions

Questions people ask

Can Kaitoi inpaint and outpaint like Invoke's canvas?

Yes, as nodes rather than layers. Inpaint, outpaint, erase and layer-splitting models run as nodes, with a drawing canvas for masks, and edit nodes change an image from a written instruction. There is no layer-based editor like Invoke's canvas.

Can I keep running models on my own GPU?

Yes. Kaitoi CLI connects your ComfyUI, Ollama and Blender, and runs Runtime recipes, on your own computer or a GPU machine elsewhere, and steps on your own hardware use no credits. Studio itself is hosted in the browser.

Is Kaitoi open source like Invoke?

Not today. Kaitoi Studio is a hosted service, and a distribution you run yourself, Kaitoi Open Source, is announced as coming soon. Invoke is open source under Apache 2.0 now.

Do I need a GPU to use Kaitoi?

No. Hosted models run on credits without your own hardware. A GPU is optional, for connecting your own ComfyUI, Ollama, Blender or Runtime recipes, and it can be in another machine from the one you use Studio on.

09 mor more comparisons

Other comparisons