Invoke 7 PR opens after 3 months and 320 PRs: new canvas engine, projects, video with audio
joshwcorbett · reddit · 2026-10-02
The InvokeAI team has opened the PR to merge Invoke 7 into the main repo — three months of work spanning 320 PRs. Highlights:
- Projects: canvas, settings, workflows, and boards live in autosaving projects
- New one-page interface: customizable widgets, instant layout switching, Ctrl/Cmd+K command palette
- New canvas engine: layer groups, adjustments, real selections, custom fonts, history panel, layered PSD export
Also included: local video generation with audio (LTX-2.5 and MiniMax H3), a gallery-wide Image Map, semantic search ("foggy forest" just works), workflows with loops, and major work on running big models in less VRAM. Not a release yet — an alpha will be tagged after merge. Requires Python 3.12; database migrations are one-way (with automatic backup). Setup instructions are in the PR for anyone comfortable building from source.
More from Multimodal
- Tavus demos fully realtime-generated AI agent and a vision of delegating identity — Kyrannio · 2026-10-02
- Boiler Room sets turn into agentic coding music hackathons as dev live-codes Temu Ryoji Ikeda — teropa · 2026-10-02
- fal launches Recast: one-click video character replacement keeping motion and audio — OdinLovis · 2026-10-02
- Fixing Krea2's blotchy noise: an algorithmic mid-frequency removal workflow with exact settings — terrariyum · 2026-10-02
- Magnific shares six Ideogram 4.5 posters with full prompt iteration threads — charis_ai · 2026-10-02
- Arctic horror short 'Variants' Episode One made with AI for under $100 — wooyi · 2026-10-02