Skip to content

Releases: LingJingAI-Labs/opencli-plugin-flow

v0.2.0 — Video editing (abra_edit)

19 May 21:26

Choose a tag to compare

What's new

🎬 Video editing mode (abra_edit)

Generate-from-video editing is now supported end-to-end:

  • flow media-upload --file ./clip.mp4 --name myclip — chunked resumable upload, sha256 dedupe, per-project cache
  • flow gen --refVideo myclip --prompt "改成晚上" — auto-detects video edit mode, 40 cr fixed
  • Video reference can be combined with image refs (--refs) to pin characters/scenes while editing

🏷️ Friendly model aliases

No more remembering `abra_t2v_8s`:

  • `--model edit`, `--model t2v-8s`, `--model r2v-4s`, etc.
  • Raw `abra_*` keys still accepted for backwards compat
  • `flow models` table now shows an Alias column

Install / upgrade

```bash
opencli plugin install github:LingJingAI-Labs/opencli-plugin-flow

or to upgrade

opencli plugin update flow
```

Internal details (for the curious)

Two non-obvious things we had to reverse-engineer:

  1. Video upload uses a 2-step protocol on `labs.google` (cookie-auth, not OAuth):
    `POST /fx/api/upload-video?action=start` → `PUT /fx/api/upload-video?action=upload` in 2 MB chunks with `X-Upload-Command: upload[, finalize]`
  2. `/v1/video:batchAsyncGenerateVideoEditVideo` rejects `useV2ModelConfig` (T2V/R2V endpoints accept it). `videoInput.{mediaId, startFrameIndex, endFrameIndex}` is how the video reference is passed.