← Apps

InvokeAI

✓ verifiedsky/invokeai

Image generation studio (canvas, gallery, node workflows), with its gallery, boards and models kept on the workspace volume.

Runs
1 GPU ≥ 24 GB · tested on L40S, A10
Image
ghcr.io/invoke-ai/invokeai:v6.14.4-cuda@sha256:315cd39086446fa81d0b2ca75615ec4407ca1ba5d1f737d825c57646608b0495
Serves
port 9090, ready at /api/v1/app/version (first start up to 30 min)
Licenses
code: Apache-2.0
Data
library/invokeai/models (assets)

Clone

Sign in to clone it

Versions

#StatusNotesContentsSealedPublishedClones
v1PublishedFirst-party app definitiondefinition only2026-11-222026-11-220

README

InvokeAI

InvokeAI is an image generation studio: a canvas, a gallery with boards, and node workflows. Deployed as a workspace, everything you make lives on a skyfs volume in your bucket and survives the GPU VM:

Stop the VM when you are done. Start a new one later, on any cloud, and InvokeAI comes back exactly as you left it.

Definition sky-app.yaml, scripts/start.sh
Image ghcr.io/invoke-ai/invokeai:v6.14.4-cuda, pinned by digest (torch 2.7.1+cu128, Python 3.12)
Port 9090, readiness /api/v1/app/version
GPUs 1 by default; --app-gpus 0 runs on the CPU
First start about 4 min on AWS (the image pull); budget 30 min

Deploy

Prerequisites: skyws worker running against Temporal, gpuprov cloud credentials and a bucket in ~/.skyws/config.toml (setup).

skyws up studio --app invokeai --cloud aws --gpus L40S:1 --region us-west-2
skyws open studio        # http://127.0.0.1:9090, over SSH; Ctrl-C to stop
skyws down studio        # InvokeAI exits, skyfs uploads the rest, the VM is terminated

InvokeAI runs in its default single-user mode, without a login, so it is never exposed on the VM's public address; skyws open is the way in.

GPU size depends on the models: 24 GB (A10, L4) runs SDXL and quantized Flux; 48 GB (L40S) runs full Flux with room for LoRAs and ControlNets.

Why upstream's image

Unlike comfyui, which builds its Python environment onto the volume over a generic torch image, this template runs InvokeAI's own image. The image is the combination upstream tests (InvokeAI pins torch 2.7 for CUDA, and its UI is a built frontend), it starts without installing anything, and the volume holds only your work and models. Upgrading is manual (docs/app-catalog.md §8): create a new workspace from a definition with a newer image digest and bring your projects across by hand. InvokeAI migrates its database on start, so a database moved to a newer version cannot go back.

The cost: packages a custom node pip installs go into the image's environment and are gone on the next VM. Nodes that need extra packages must be reinstalled after each start, or the image extended.

What is kept, and where

The app container sees the volume at /sky/volume:

Path What Kept
projects/<workspace>/ InvokeAI's root: invokeai.yaml, databases/ (gallery, boards, model records, workflows), outputs/, nodes/, style_presets/, workflow_thumbnails/ yes
library/invokeai/models/ installed models, shared by every InvokeAI project on the volume yes
library/hf/, home/ HF_HOME, HOME yes
/cache/invokeai/downloads partial model downloads, on the VM no

A model installed from one project is on the shared disk but recorded in that project's database; another project adds it with Scan Folder on library/invokeai/models, installed in place.

Configuration

start.sh sets what the workspace depends on as INVOKEAI_* variables, which take precedence over invokeai.yaml: host, port, the models directory, the download cache and, without a GPU, the device. InvokeAI writes invokeai.yaml into the project on its first start; edit it there for everything else (reference), e.g. max_cache_vram_gb, precision, remote_api_tokens. Restart InvokeAI (skyws up again) to apply.

Durability

skyws down stops InvokeAI first, so its SQLite database is closed, then waits for skyfs to upload the rest and write a snapshot. If a VM disappears without a down, the volume keeps everything up to about a second before.

Publishing a workspace as a template version includes its gallery: the database refers to the images in outputs/, so they are kept together.

Troubleshooting

gpuprov ssh studio -- sudo k3s kubectl -n sky logs deploy/studio -c app     # InvokeAI
gpuprov ssh studio -- sudo k3s kubectl -n sky logs deploy/studio -c skyfs   # the volume

Verified

AWS g6e.xlarge (NVIDIA L40S) in us-west-2, S3 bucket, provisioned by the workflow:

Not yet verified: a generation, a model install, skyws down and a restart onto a new VM.

Scripts

What runs in your workspace, from definition sha256:0c337a09669e604e2368d4ad1613dbaafff587fce69b247ff50478c500eee370.

scripts/start.sh
#!/usr/bin/env bash
set -euo pipefail

# The project is InvokeAI's root: invokeai.yaml, the database (gallery,
# boards, model records, workflows), outputs, custom nodes and style
# presets. Models are shared by every InvokeAI project on the volume.
project=$SKY_ROOT/projects/$SKY_WORKSPACE
models=$SKY_ROOT/library/invokeai/models
mkdir -p "$project" "$models" "$HOME" /cache/invokeai/downloads

# What the workspace needs fixed is set here: INVOKEAI_* variables override
# invokeai.yaml, which InvokeAI writes on the first start and is the user's
# to edit for everything else (memory, precision, logging, API tokens).
export INVOKEAI_HOST=0.0.0.0
export INVOKEAI_PORT=$APP_PORT
export INVOKEAI_MODELS_DIR=$models
# Partial downloads stay on the VM; only installed models reach the volume.
export INVOKEAI_DOWNLOAD_CACHE_DIR=/cache/invokeai/downloads
if [ "${SKY_GPUS:-1}" = 0 ]; then
  export INVOKEAI_DEVICE=cpu
fi

cd "$project"
exec invokeai-web --root "$project"