Skip to content

AI Image Workbench 0.9.8

Choose a tag to compare

View all tags
@margetrp-hub margetrp-hub released this 06 Jun 05:43
· 34 commits to main since this release
v0.9.8
823d327

0.9.8

This release is the first step toward a real 1.0 beta, but it is intentionally not labeled 1.0 yet.

AI Image Workbench is moving from an early gateway-specific starter into a standalone image creation workstation that can run with an OpenAI-compatible gateway, while still keeping backward compatibility with existing Sub2API deployments.

What Changed

  • Public package name changed to ai-image-workbench.
  • Front-end configuration now prefers generic gateway variables:
    • VITE_AI_GATEWAY_BASE_URL
    • VITE_AI_GATEWAY_MODEL_BASE_URL
    • VITE_AI_IMAGE_ROUTE
    • VITE_AI_RESPONSES_MODEL
    • VITE_AI_GATEWAY_LOGIN_URL
  • Docker and Nginx proxy configuration now prefer AI_GATEWAY_UPSTREAM.
  • The history/session service now prefers AI_GATEWAY_BASE_URL.
  • Existing VITE_SUB2API_*, SUB2API_UPSTREAM, and SUB2API_BASE_URL variables remain as compatibility aliases.
  • Docker Compose defaults to STUDIO_AUTH_MODE=local, so the workbench can persist sessions/history without depending on an upstream account system.
  • Gateway-authenticated deployments can still use STUDIO_AUTH_MODE=gateway.
  • Provider settings now separate credential source (apiKeySource) from provider family (providerId), so existing gateway accounts, manual OpenAI-compatible APIs, and future NewAPI-style adapters have a clearer path.
  • src/studio/providers/registry.js now stores route, auth, capability, parameter, and default-model metadata for provider families.
  • Server-side generation jobs now persist provider metadata and normalize orphaned active jobs to unknown after a service restart or lost runner, instead of leaving them looking actively queued forever.
  • Browser-side history now uses IndexedDB as an expanded local cache with localStorage as a fallback, reducing the chance that larger local histories disappear or overload localStorage.
  • The history gallery now renders local session cards in batches, reducing the initial DOM and image-node pressure when a user has many saved sessions.
  • The video inspiration gallery now renders cards in batches as well, keeping larger idea libraries lighter in the browser.
  • Image template category/search results now render cards in batches too, so large prompt libraries avoid mounting every card at once.
  • JSON persistence reads now tolerate UTF-8 BOMs, which makes Windows-authored backups and manual VPS recovery files safer to load.
  • Large-canvas performance mode virtualizes offscreen image/video nodes and reduces SVG line animation load.
  • Manual provider API keys are session-only in the browser. Provider configuration can persist, but raw manually entered API keys are migrated out of localStorage and kept only in sessionStorage for the current browser session.
  • The local release gate now includes documentation encoding checks, so broken README or docs mojibake is caught before publishing.

Why This Is Not 1.0 Yet

The workbench now has a cleaner standalone deployment shape and generic gateway naming, but the provider abstraction is still transitional. A true 1.0 should finish the provider registry, make model capabilities explicit per provider, and document a stable extension path for OpenAI, NewAPI, Sub2API, and other compatible gateways.

Assets 6
Loading