InvokeAI Local Install: Image, Canvas & Video

InvokeAI is the local install for canvas-first image work, with 6.14.0 adding short Wan 2.2 videos through workflows. Generation is private by default because it runs on your hardware.

At a Glance

TL;DR — Is It Worth It?

Use InvokeAI when you want a polished local canvas: paint, mask, extend, and composite without building a node graph first. That workflow has not changed in 6.14.

What did change: you can now generate short local videos with Wan 2.2. Official docs put that path in the workflow editor, with starter templates for text-to-video, image-to-video, and first-last-frame interpolation. Upstream still marks video as experimental.

The tradeoff versus ComfyUI and Forge is the same as before. Those UIs are usually faster for raw batch image work, and ComfyUI still tends to get brand-new architectures earlier. InvokeAI follows when the canvas and model manager are ready. If you only want the newest model the day it drops, start with ComfyUI.

What's in InvokeAI 6.14.0

Verified against the 6.14.0 release notes:

• Local Wan 2.2 video — text-to-video, image-to-video, and image-to-image-video (interpolate between two stills). Template workflows can concatenate clips into longer sequences.
• Current image models — Krea.2 Turbo and Raw, Flux.2 Dev, Ernie Turbo, Ideogram 4, plus Anima ControlNets and inpainting.
• Flux.2 PiD — super-resolution up to 4K.
• Multi-GPU — queued jobs can run in parallel across installed GPUs; a single job can split the text encoder and denoiser.
• Native Intel XPU — official hardware notes now list Intel Arc (Alchemist, Battlemage and newer) on Windows and Linux x86_64, with the same VRAM bands as Nvidia.
• FP8 storage — opt-in in the Model Manager for selected families (Anima, Flux.2 Krea Turbo and Raw, Z-Image Turbo). Upstream says this can cut resident model size in VRAM by about half.

Some of those new image families are text-to-image only, or have staged LoRA / ControlNet support. Do not assume every 6.14 model has full canvas inpaint, regional guidance, or reference-image wiring. The release notes include a capability table.

For video VRAM, 6.14 documents an opt-in: add wan_memory_optimization: true to invokeai.yaml to reduce decode pressure at some speed cost.

Top Features

1. Unified Canvas — txt2img, img2img, inpainting, and outpainting in one infinite workspace. Layers, brushes, masks, and regional prompts. This is still the product's best local feature.

2. Workflow editor (now including video) — node graphs for custom pipelines. In 6.14 the same editor is the supported Wan 2.2 video path. It is simpler than ComfyUI, and that is the point.

3. Model manager — download and switch checkpoints, Diffusers trees, LoRAs, and embeddings from the UI. 6.14 adds starter bundles for Wan 2.2 and the newer image families.

4. Gallery and boards — organize outputs, recall metadata, and remix a past image with the settings that made it.

5. ControlNet and regional prompting — guidance layers on the canvas for families that support them. Check the 6.14 matrix before you expect this on Krea.2, Ideogram 4, or Ernie Turbo.

Local Install Path

Official install is the Invoke Launcher: download the launcher for Windows, macOS, or Linux, click Install, and let it manage the Python environment. Manual uv and Docker paths exist if you do not want the launcher.

The local WebUI defaults to port 9090 (http://localhost:9090). First launch still needs network access to pull models. After that, generation can stay on-device.

This page is about that local/private install. Upstream currently lists third-party hosted GPU options (AI Badgr, RunPod, Railway) if you do not want to use your own card. Those hosts are not Invoke Studio, and they are not the default path this review is for.

Requirements (official guidelines)

Hardware notes are rough best-performance bands. Cards with less VRAM often still run, slower. Quantized / FP8 builds change the floor.

Other official platform notes:

• Apple Silicon Macs work; 16 GB+ unified memory is recommended. The official download page says macOS is Apple Silicon only.
• AMD GPUs are documented as Linux / ROCm. Treat Windows AMD as unverified here.
• Intel Arc is listed for Windows and Linux x86_64 with the same VRAM bands as Nvidia.
• Flux.2 Dev is in the 6.14 feature list. Upstream does not publish a Dev row on that hardware table, so this page does not invent a VRAM number for it.

Limitations

• Video is experimental and workflow-only. Do not expect Wan 2.2 on the Unified Canvas. Official docs describe short clips (about 81 frames / 5 seconds at 16 fps) and concatenation for longer pieces.
• New-model lag. ComfyUI still usually ships day-one graphs. 6.14 closed a lot of that gap for Wan 2.2, Flux.2 Dev, and Krea.2, but it is still a follow-on UI.
• VRAM residency. InvokeAI can hold GPU memory until you restart or clear the cache. invokeai.yaml and the Model Manager FP8 toggle are the documented levers.
• Shallower node graph than ComfyUI. Fine for canvas work and the shipped video templates. Weak if you live in huge custom graphs.

How It Compares

Bottom Line

Use InvokeAI if:

• You do iterative canvas work — faces, backgrounds, composites — and want one local workspace.
• You want a local install with a model manager, boards, and regional prompts built in.
• You want short Wan 2.2 video on the same box, and you are willing to use the workflow templates.

Skip InvokeAI if:

• Raw image speed is the only metric. Forge and ComfyUI still win that comparison.
• You need every new architecture on day one. Start with ComfyUI.
• You want video as a canvas brush. That is not what 6.14 shipped.

If you want a hosted GPU instead of a local card, use the third-party hosts listed on invoke.ai. This review stays on the private local install.

For a zero-setup option that gets you generating immediately, LocalForge AI ships with Forge pre-configured and ready to go - no Python environments, no YAML files, no troubleshooting. It's one path among several, but it's the fastest way from download to first image.