Send images to InvokeAI's image or video generation tab

[lstein]

Summary

The metadata drawer's InvokeAI section now lets you choose which InvokeAI tab Send Image and Append Image feed:

InvokeAI
[Send Image] [Append Image] as a reference for <image generation | video generation>
[Remix] [Recall]

• image generation (the default) works exactly as before.
• video generation uploads the image (or reuses InvokeAI's copy) and places it through InvokeAI's new POST /api/v1/recall/video/{queue_id}/image route (feat(video): place a gallery image in the Video panel from the recall API invoke-ai/InvokeAI#9693). The Video panel's model decides where it lands:
  • Models with reference images (MiniMax H3 Ref2VA): Send replaces the reference images, and Append adds one.
  • Frame models (Wan 2.2 I2V, LTX-2): Send sets the first frame, and Append fills the next empty frame. Further appends are ignored.
• InvokeAI reports the outcome in a notice in its own window.

Details

• Gating: /invokeai/capabilities reports a new video_image flag, probed from the backend's /openapi.json. The pulldown is hidden, and Send/Append always go to image generation, until the configured InvokeAI has the route.
• Append in video mode: Append Image is shown whenever video generation is chosen, even on backends whose image recall lacks append.
• Persistence: the choice is a server-backed per-device preference (invokeRefTarget in PERSISTED_SETTINGS, invoke_ref_target in UserPreferences). It therefore survives iOS storage purges and is cleared by Reset to Defaults.
• Where the logic lives: invoke-ref-target.js owns the pulldown. Both drawer renderers (swiper and grid) re-apply it after each re-render.
• Videos are unaffected: video files never get the pulldown, and /use_ref_image still refuses them.
• Layout: the row wraps cleanly in a narrow drawer.
• Docs: docs/user-guide/invokeai-integration.md documents the choice and the per-model placement rules.

Testing

• Backend: 1306 passed (full suite before the rebase; the touched files were re-run after it). New coverage:
  • use_ref_image with target="video": upload, then the /image route with image_name/append and an empty body; an upstream refusal returns 502; an unknown target returns 422.
  • The video_image capability flag.
  • Formatter rows and pulldown placement.
• Frontend: 1019 Jest tests pass. New coverage:
  • invoke-ref-target.test.js
  • target selection in invoke-recall.test.js, including a remembered "video" being ignored on an unsupported backend
  • the capability class
• ruff, eslint and prettier are clean.
• Manually tested end to end against InvokeAI running feat(video): place a gallery image in the Video panel from the recall API invoke-ai/InvokeAI#9693.

🤖 Generated with Claude Code