InvokeAI 7 Alpha 3: Board Sharing Limits, Qwen and Upgrade Risks

InvokeAI 7 alpha 3 introduces a board-role model, adds Qwen-Image-2.1 support and makes SageAttention optional for supported GPUs. Board sharing currently uses general-access settings; individual role assignment is not yet available in the interface. Test these changes separately from your working v6 installation.

The October 19 release is a prerelease. It also changes the database and project format, so a quick upgrade followed by a downgrade is not a safe testing plan.

Start with a separate, empty root directory

Quit InvokeAI and back up your existing root directory before experimenting. Include its configuration, database and outputs, and check whether your configuration stores any of those elsewhere. Keep the v6 installation and its original root unchanged.

In Invoke Launcher, choose a new, empty install location. Select Manual when choosing a version and enter v7.0.0-alpha.3. A manual installation can use --root /path/to/new/folder or INVOKEAI_ROOT to select the separate root.

Pointing v7 at the v6 root triggers a one-way database migration. v6 cannot open that upgraded database. Invoke makes a backup before migrating, but restoring it returns to the earlier state; work created in v7 will not appear in the restored v6 gallery.

If you have already migrated accidentally, follow the recovery procedure in the release notes. It covers the database backup, SQLite sidecar files and configuration changes that a simple reinstall would miss. Keep the database files while recovering.

Alpha 3 also saves projects in a format alpha 2 cannot open. Preserve earlier project copies before testing exports, and keep production work in the version you already trust.

Which board-sharing controls can you use?

The tagged multi-user guide describes Viewer, Contributor, Editor and Owner roles, but also says owners cannot yet assign a role to one user at a time. The available sharing control is general access: Restricted, Everyone can view or Everyone can contribute.

Viewer access permits browsing, downloading, copying and using board media in generations. Contributor access also lets people add media and manage their own additions. That gives a team a way to share a reference library or collect outputs, provided the same general-access level suits the other accounts on the installation.

For an evaluation, create a disposable board and switch its general-access setting. Check visibility and contributions from separate non-administrator accounts. An Editor role exists in the permission model, but do not plan a workflow around assigning it to a specific teammate through the current interface.

The database migration maps old private boards to restricted access, shared boards to viewer access and public boards to contributor access. Review those outcomes when moving a shared installation, especially if a board contains work intended for a small group.

Project settings are also separated by member, including layouts, generation settings and prompt history. That is useful groundwork for collaboration, although the release describes broader project collaboration as future 7.X work. Board sharing should be evaluated on the controls available today.

How complete is Qwen-Image-2.1 support?

Both the Base model and eight-step Turbo variant work with text-to-image and Canvas image-to-image, inpainting and outpainting. Invoke's alpha 3 model guide also lists important gaps: reference-image editing, LoRAs and regional guidance are not supported yet.

Use the starter models for a first test. A standalone GGUF or single-file transformer still needs the Qwen-Image-2.1 VAE and a Qwen3-VL 8B encoder selected in Components. The older Qwen Image VAE and Krea-2's 4B encoder are not interchangeable with them.

Base defaults to 40 steps; Turbo defaults to eight. Compare both on your actual prompts and inspect text accuracy, unwanted changes and transparent edges. Fewer steps can shorten iteration, but the result still needs to survive whatever editing or compositing comes next.

Transparent output has a specific limit here. The VAE can generate an alpha channel, while Canvas image-to-image, inpainting and outpainting treat the input as opaque. Include a transparent asset in your evaluation if preserving its existing alpha matters to your workflow.

The weights also carry the Qwen Research License, which limits the granted use to research or evaluation. Commercial use requires a separate license from Tongyi. Check that permission before planning client work or a product around these weights; their availability in Invoke does not change their license.

When does SageAttention help?

SageAttention is an optional attention backend for supported diffusion models. It needs SageAttention 2.2 installed separately and a supported NVIDIA GPU, such as an RTX 30-series card; Jetson Orin is excluded. Invoke's versioned setup and benchmark guide covers the dependencies. Enable it with attention_backend: sage in invokeai.yaml, restart, and check the startup log.

Invoke's maintainers measured these gains on an RTX 4090 on Windows with models already loaded. FLUX.1 dev in FP8 at 2048×2048 fell from 84.2 to 54.8 seconds, while its 1024×1024 result went from 16.1 to 14.0 seconds. Those figures describe particular workloads; Linux gains were not measured.

Memory pressure can reverse the benefit. For Wan 2.2 T2V A14B in GGUF Q4_K_M with a four-step Lightning LoRA, denoising an 81-frame clip at 1280×720 on a 24 GB GPU increased from 293 to 662 seconds. The extra memory required by SageAttention matters when a generation already nearly fills the card.

The quantized kernel can also change the image at the same seed. Compare output quality as well as elapsed time, and use attention_backend: auto for the baseline. Record the backend beside your test results because image metadata does not preserve that setting.

What should pass before you consider switching?

Use a small evaluation that reflects the work you need to finish. Test the available board general-access settings, run a familiar Canvas edit and reopen a project after saving. If you use SageAttention, measure a normal job and one near your memory limit, with the same model and resolution in each comparison.

Keep failures and settings together so you can reproduce them. The useful outcome is knowing whether a specific workflow works reliably enough for your needs. Until it does, the separate alpha root lets you continue evaluating without making your working v6 database part of the experiment.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn technical launches into clear product stories and help founders get those stories in front of the people who will use what they build.

Book a 30-minute call.