Reconstruct An Edit From A Reference
You have a finished or approved reference video, but the timeline behind it is missing. Perhaps a trailer was cut from watermarked proxies, an offline editor sent only a render, or the online edit must be rebuilt from clean camera originals. The reference shows the correct shots and edit points; your task is to find those moments in the available source media and reconstruct them as editable Premiere clips.
The library workflow Reconstruct Edit From Reference performs that first conform pass. It searches the original media you select in the Project panel, identifies the most likely source interval for every reference shot, progressively fills empty shot ranges on a target video track, leaves uncertain results reviewable, and writes a durable analysis report. Existing target clips are always preserved.
This recipe uses Automation Agent for Adobe Premiere. Start with Set Up Your AI if the panel is not installed yet. The ready-made workflow itself uses deterministic image analysis and does not require a Workflow AI provider.
The Complete Reference-Conform Workflow
The finishing tools are separate because each also solves a useful problem on its own. For a complete reconstruction, use only the stages your material needs:
| Stage | Tool | Use it when |
|---|---|---|
| 0. Pre-index reusable sources | Build A Reusable Auto-Conform Source Index | The same feature, episode, or campaign footage will be searched for several different reference edits. |
| 1. Find shot boundaries | Detect Scene Edits In Selected Clips or the preparation option inside Reconstruct Edit From Reference | The reference is one flat rendered clip rather than already split at every cut. |
| 2. Find and place source shots | Reconstruct Edit From Reference | The original clips are not yet on the target track, or you do not know which source intervals were used. |
| 3. Refine source timing | Slip Clips To Match Reference | The correct source clips are present, but their visible frames do not align exactly with the reference. |
| 4. Recover crop and placement | Match Clips To Reference Framing | The edit timing is correct, but Position and Scale do not reproduce the approved framing. |
Reconstruction often finds the correct frame already. Slip matching remains useful for manually assembled rough conforms, difficult matches, or a separate verification pass. Framing recovery is optional and independent of source-shot retrieval.
Fastest Path: Run The Library Script
- Install the FFmpeg and Python/OpenCV Runtime Packs.
- Put the reference on one video track in the active sequence. The default is V2.
- Choose the target video track. The default is V1. It may already contain a partial conform: the workflow fills only reference-shot ranges that are completely empty.
- In the Project panel, select every original source clip that the workflow should search. This explicit selection keeps both runtime and false matches bounded.
- If these sources were pre-indexed, choose the same local Auto-Conform workspace used by Build Auto-Conform Source Index.
- Open
Library > Examples > Finishingand run Reconstruct Edit From Reference.
No live agent session or prompt is required. The matching, confidence policy, placement, markers, cleanup, and report format are frozen into the reviewed library workflow.
The workflow never clears, trims, or replaces an existing target clip. Any positive overlap protects that complete reference-shot range; a clip that only touches the shot boundary does not block it. This also applies to red low-confidence fallback clips. Delete a wrong or red fallback clip and run the workflow again to make that shot eligible. If a run is canceled or fails, rerunning safely preserves earlier placements and continues with the remaining gaps.
Reference Preparation
Choose one of two modes when the workflow starts:
- Already cut reference (Recommended) expects a separate reference timeline clip for every shot. Use this when Scene Edit Detection has already been run or the reference was assembled from individual clips.
- Flat reference: detect scene edits first expects exactly one reference clip and runs Premiere's native Scene Edit Detection before matching. The flat clip is split in the timeline, then the workflow relists the newly created segments and uses them as reference shots.
If you want markers or Project-panel subclips instead of timeline cuts, run the separate Scene Edit Detection recipe.
Confidence, Placement, And Review
For each reference shot, the matcher samples the rendered reference, searches cached descriptors from the selected sources, verifies multiple frames, and refines the best candidate to a frame-aligned source interval. Complete analysis batches are applied as they finish, so clips appear progressively during a long run. Immediately before each placement, the workflow relists the target track and skips the shot if its range has become occupied.
The default policy, Place the best available candidate for every shot and flag uncertain results, gives you a concrete result to inspect whenever the matcher has one structurally usable candidate. Confidence controls the marker and review priority rather than suppressing that result:
- Green markers identify automatically confirmed matches that were placed.
- Yellow markers identify plausible but ambiguous matches. They are placed by default so you can inspect the algorithm's best result directly in context.
- Red markers identify results the algorithm does not trust. A structurally usable best candidate is placed as a red fallback by default; failures and shots without any usable candidate remain gaps. Choose Leave low-confidence results as gaps for a conservative conform that does not place yellow or red candidates.
Keep Place ranked alternatives for uncertain shots enabled when you want a ready-made review stack. For each yellow or red shot, every other retained, fully refined candidate is mapped by rank to an existing video track above both the reference and winner tracks: rank 2 uses the first such track, rank 3 the next, and so on. Add enough empty video tracks above both primary tracks before starting if you want every retained rank to appear. Alternatives are placed video-only, disabled, and independently verified, so the winner remains visible. The workflow never creates tracks and never moves a rank to another lane: a missing lane or any foreign clip overlap leaves that alternative unplaced and records the reason instead of overwriting editorial work. A failed optional alternative is reported without aborting the already reconstructed winner conform after any exact fresh enabled placement has been removed safely. If Premiere also rejects that cleanup and an enabled higher-track clip could cover the winner, the workflow fails instead of reporting an unsafe partial result. Green auto-confirmed shots do not receive alternative stacks.
A red fallback is deliberately only a lead. Titles, graphics, unfinished or missing VFX shots, and source media omitted from the Project-panel selection may have no correct candidate at all, yet the matcher can still show its best wrong result. This makes the available evidence reviewable instead of forcing an immediate search through all footage, but it does not turn NO_RELIABLE_MATCH into a trusted match.
Winner marker comments include the shot, source, confidence score, runner-up separation, and placement outcome. A placed red fallback uses the explicit PLACED_UNRELIABLE marker outcome. Alternative point markers preserve rank, fixed track, source and time identity, score, and placement status. A rerun reuses an old alternative only when that marker identity and the disabled timeline clip still match exactly; changed or unrelated clips are preserved as occupied content. If an earlier interrupted run left an exact workflow-owned alternative enabled, the rerun repairs that item in its fixed lane instead of treating it as foreign editorial content. Markers are written only after fresh checks confirm the project item, sequence boundaries, source in/out, and—for alternatives—the disabled state of every new placement. The selected Auto-Conform workspace receives a JSON winner result, a human-readable .txt processing-time report, a separate ranked-alternative TSV, and an append-only placement journal under reports/. The text report separates source indexing, shot analysis, timeline placement/final verification, and alternative placement outcomes, so production timings can be shared without reading JSON. Reusable source indexes remain under cache/, so later runs with the same media do not need to decode the full sources again. Uniquely named temporary work belongs under runs/ and is removed after a successful run unless debug artifacts are requested.
Expected Result In Premiere
Newly filled target ranges contain editable video clips with timeline boundaries matching the reference shots and explicit source in/out ranges from the selected originals. Previously occupied winner and alternative ranges remain untouched. No audio is placed. Plausible but uncertain winners and structurally usable red fallbacks are placed and visibly marked by default. Their safe lower-ranked candidates appear disabled on deterministic higher tracks when those tracks exist and are empty. Shots without any usable winner remain gaps; the conservative policy also leaves yellow and red winners as gaps.
Review the conform in this order:
- Review every red fallback or gap, then inspect yellow-marked placements.
- Compare source timing against the overlapping reference. Run Slip Clips To Match Reference if a clip needs frame-level adjustment.
- Compare crop and composition. Run Match Clips To Reference Framing when Position or Scale must be recovered.
Limitations And Performance
- Reference segments and source candidates must be direct, online media files readable by FFmpeg. Nested sequences, generated media, offline clips, and ambiguous multicam interpretations are not supported.
- Every reference segment must come from one direct media file. Timeline-only effects are visible only if they are baked into a rendered reference.
- Speed changes, reverse playback, time remapping, dissolves, and other non-hard transitions are not reconstructed.
- The workflow reconstructs video only. Audio conform is not yet included.
- Heavy crop or reframing can reduce retrieval confidence. Use the dedicated framing workflow after the correct source shots are present.
- Long sources can take several minutes to index on the first run. Cached repeat runs are much faster, but frame-level verification still performs random-access decoding in the source media.