LoginSign Up
← Back to all sparks
I

InvokeAI

AI-ASSISTANTS
Velocity6.3

Open-source generative AI image creation and workflow studio

invoke.com ↗

InvokeAI 7 alpha 3 turns boards into shared, role-gated spaces and adds Qwen-Image-2.1.

local-aiimage-generationmulti-userpermissionsui-rewriteopen-source
◆Current state
InvokeAI is a local-first image and video studio three alphas into a v7 rewrite. Alpha 1 replaced the tabbed UI with autosaving projects and a widget workbench; alpha 2 added MySQL/MariaDB backends and per-project boards. Alpha 3 adds board sharing with Viewer, Contributor, Editor and Owner roles, keeps each member's settings separate inside a project, and brings back a long list of legacy features. It also adds Qwen-Image-2.1 and an opt-in SageAttention backend that cuts large NVIDIA generations by 15–35%.
◆Where it's heading
Each alpha moves InvokeAI further from a single-user desktop app toward a shared workspace: pluggable databases, then project boards, now per-role permissions and per-member settings that the notes call groundwork for collaboration later in 7.x. Model support keeps pace in parallel, and the priority is running new models on existing hardware through GGUF, fp8 and quantized attention. Restoring legacy features signals the team wants v7 at parity before it asks v6 users to migrate.
◆Prediction
Expect alpha 4 within about a week, with more parity restoration and collaboration plumbing; project file formats may break again, since alpha 3 already did. The entries still give no beta or stable date.

◆Recent moves

  1. 10d ago

    InvokeAI 7 alpha 3 adds role-based board sharing and Qwen-Image-2.1

    Alpha 3 continues the shift toward shared installations: boards get owners and four access roles, and projects separate each member's settings from shared content. It also adds Qwen-Image-2.1, an opt-in SageAttention backend, and restores legacy features such as the SDXL Refiner and queue retries. Project files saved here won't open in alpha 2.

    View source ↗
  2. 18d ago

    InvokeAI 7.0.0 (alpha 2)

    Alpha 2 builds out the project model introduced in alpha 1: projects can now hold unlimited boards and be moved as a whole. It also adds MySQL and MariaDB as alternatives to SQLite for multi-user installs. Most of the rest is GGUF loaders for Krea-2, Ideogram 4 and Ernie, a reserved_vram_gb setting for out-of-memory errors, and model-manager fixes, which is the expected stabilization of the v7 rewrite.

    View source ↗
  3. 24d ago

    InvokeAI 7 alpha: autosaving projects and a single widget workbench

    ⚡ SPARK

    After a 6.x line spent adding models, v7 changes how the app is organized: projects carry their canvas, settings, workflows and queue history, and the separate tabs become one dockable workbench. It is the first release since 6.14 that changes the product's shape instead of its model list.

    View source ↗
  4. 1mo ago

    6.14.2 adds custom Canvas fonts and Krea-2 style transfer

    A maintenance release that also adds imported fonts for the Canvas text tool and reference-image style transfer for Krea-2. It includes a path-traversal fix in serializer names and model keys, continuing the security cleanup from 6.14.1.

    View source ↗
  5. 1mo ago

    Web v2 can find a conditioning image in the gallery from its thumbnail

    A build-tag merge that brings a small gallery feature to the new web front end. It is an early sign of the v2 interface work that shipped as v7.

    View source ↗
  6. 1mo ago

    6.14.1 adds workflow PNG export and DoRA LoRA support

    A follow-up to 6.14 that adds workflow screenshots saved as PNG, middle-click to open media in a new tab, and Krea-2 LoRAs with DoRA magnitudes. It also closes a security hole, a pattern that continues through the 6.14 patch line.

    View source ↗