Send videos to InvokeAI 7: Initial Video, Ref Video, Recall and Remix

[lstein]

Summary

InvokeAI 7 adds a video recall API (invoke-ai/InvokeAI-7#339) that mirrors the image recall API PhotoMap already uses. This PR adds InvokeAI buttons to the metadata drawer for videos:

Button	Shown for	What it does
Initial Video	every video	Puts it in the Video panel's Initial Video slot (the clip that gets extended)
Ref Video	every video	Appends it to the reference videos of the selected model (MiniMax H3 Ref2VA today)
Remix	videos with InvokeAI video metadata	Recalls every parameter except the seed (mode=remix, strict=true)
Recall	videos with InvokeAI video metadata	Recalls every parameter, including the seed (mode=recall, strict=true)

The buttons only appear when the linked InvokeAI's OpenAPI schema lists /api/v1/recall/video/{queue_id} (the new video_recall capability). Image buttons are unchanged.

How it works

• Record → request (metadata_modules/invoke/video_recall.py):
  • The request is built from a whitelist over the raw record, not the Pydantic parse. The upstream request model is extra="forbid", and the raw record also carries ltx2_* fields PhotoMap doesn't model.
  • Models are sent by name, and LoRAs as model_name.
  • Each value is checked against the upstream bounds and dropped on its own if it fails. Upstream rejects the whole request (422) over a single bad value.
• Sending the file (POST /invokeai/video/use_media):
  • A video InvokeAI may already have (UUID filename, or an InvokeAI record) is placed by name first.
  • On 401/403/404 it's uploaded through the /upload route instead, onto the configured board, falling back to Uncategorized.
  • A timeout while InvokeAI converts the video gives a message saying the video may still appear, so the user doesn't retry and create a duplicate.
• Recall (POST /invokeai/video/recall):
  • Parameters InvokeAI couldn't use (listed in its skipped) are shown in a yellow note under the buttons.
  • If InvokeAI refuses the recall with a 403 because the record names media the user can't read, it's retried once without the media fields, and those are reported as skipped.

Also in this PR

• Auth fallback fix (invokeai_client._request_with_auth_fallback). A 403 on a request that carried a token was always retried anonymously, and the token dropped. When that anonymous retry is refused with a 401, the backend still requires login, so the 403 was a real permission refusal: it's now returned as-is and the token kept. Before this, a board the user couldn't write to broke the board fallback on every other click, showing "401 Not authenticated". This affects image uploads too.
• Test isolation fix (tests/backend/conftest.py). PHOTOMAP_CONFIG was set in a session fixture, which runs after test collection. A test module that imports a router at module level would bind the cached config manager to the developer's real config file, and the tests would then write to it. This happened while developing this PR. The variable is now set when conftest is imported, and a session check fails the run if the config path isn't the temp file.

Known limitations

• The upstream API (feat(api): external video recall API invoke-ai/InvokeAI-7#339) isn't merged yet, and the InvokeAI frontend handler for the video_recall_requested event ships in a follow-up PR. Until that lands, the buttons report success but the InvokeAI UI doesn't change.
• LTX-2 conditioning videos go through the same Ref Video button, so InvokeAI decides where a reference video goes based on the selected model. The API has no append route for the single LTX conditioning slot.
• Models are sent by name, not by key, the same as image recall, so recall also works on a different InvokeAI install. On the install that produced the video, two models with the same name could resolve to the wrong one.

Test plan

• uv run pytest tests: 1171 passed. The new tests/backend/test_video_recall.py covers the translation, both routes, board and auth fallbacks, withheld media, the timeout message and the capability flag.
• npm test: 915 passed. New cases cover the four video modes, the skipped/warning note, success: false, and the capability class.
• ruff check, npm run lint, npm run format:check
• An adversarial fresh-context review; its confirmed findings are fixed here.
• Manual: against a backend running InvokeAI-7#339, check that Initial Video, Ref Video, Recall and Remix send correctly and that InvokeAI emits video_recall_requested.
• Manual: against an older InvokeAI, the video buttons are hidden.

🤖 Generated with Claude Code