Choose one edit that serves the scene
A useful video edit begins with a specific production problem. In a drama insert, the notebook on a desk may need to become a different prop while the lamp, wall and camera framing remain acceptable. That is a narrower brief than generating the entire scene again. Write both the requested change and the details that must survive before preparing a mask.
The official VACE repository describes reference-to-video, video-to-video and masked video-to-video tasks. Its published Wan models are identified as Wan2.1-VACE. Record the exact model and implementation instead of treating every similarly named community workflow as the same release. This page covers the official-project input and review process; the Wan overview remains the broader model-family guide.
| Starting job | Preparation question | Acceptance question |
|---|---|---|
| Reference-led generation | Which appearance reference is approved | Does the returned subject match the brief |
| Global video edit | Which source movement and scene structure matter | Does the complete shot retain the intended action |
| Masked local edit | Which region may be regenerated in every frame | Are target, edges and surrounding scene acceptable |
Separate the source mask and missing region
The official input guide defines src_video, src_mask, optional reference images and a descriptive output prompt. White mask areas identify generated regions; black identifies retained regions. In its conditioning representation, missing video areas use gray value 127. The mask sequence must match the video's shape. Those are separate jobs: a mask identifies the editable region, while the conditioning video supplies retained context.
Do not assume that a white mask erases an object by itself. It marks a region for generation. Your output description and the model's returned content still need review. A region that is too small can leave pieces of the old prop; one that is unnecessarily large exposes more surrounding scene to regeneration.
Inspect the original aligned mask exercise
Our original authored exercise moves a red notebook to the right across a simple desk scene. A rectangular white mask follows it with a margin. A separate conditioning sequence places gray 127 in that region. This is a geometry and file-alignment study, not a VACE output, tracked real footage or proof that a checkpoint accepts these files.
Watch this clip and view its details
Open the conditioning study watch page. The silent browser preview places source, mask and conditioning side by side. Download the separate lossless masters for inspection: original source, matching mask and conditioning sequence. MKV masters use FFV1 for lossless inspection and may need a desktop player or conversion appropriate to your selected workflow; they are not a universal model-ready delivery format.
Each master has 81 frames at 16 fps, 640 by 360 pixels and an actual duration of 5.0625 seconds. The frame geometry ledger declares each rectangle. We decoded all 81 frames of each master: mask pixels matched the declared black-and-white geometry, conditioning pixels inside the rectangle were 127 and pixels outside it matched the source. The scoped verification report includes the file hashes. These checks establish this authored exercise's pixels, not model inference or subject tracking on arbitrary footage.
Check the pair before diagnosing generation
Download the free read-only source and mask checker. It uses an existing local ffprobe to count decoded visual frames and read dimensions and frame-rate metadata. It installs nothing, writes no input files and restricts media access to local files.
python3 check-vace-pair.py --source source-video.mkv --mask edit-mask.mkvThe checker requires matching known dimensions, decoded frame counts, average frame rate and nominal frame rate. Unknown or unreadable values fail conservatively. It does not inspect mask pixels, verify per-frame timestamps, find the subject or certify a model-ready input. A pair can pass container alignment and still have the wrong polarity or a mask that loses the prop during motion.
Review the first, middle and last frame, then follow the moving edge through the full sequence. For actual footage, check turns, occlusion and motion blur. If another object passes in front of the target, decide explicitly whether that foreground region is editable or retained. Do not let a bounding rectangle make that story decision accidentally. The rotoscoping workflow covers more detailed boundary review.
Run the preprocessing path for your implementation
Prepare inputs using the current instructions for your selected task. The official project separates preprocessing from inference and also supplies an end-to-end pipeline. Its Wan examples show source video, mask, reference and prompt arguments. We did not install model weights or execute those inference commands for this article.
In a ComfyUI integration, inspect the expected file types, node versions, mask polarity, resize behavior and frame handling. Save the graph alongside your accepted take. A community graph labeled Wan VACE or Wan VACE GGUF is a specific integration, not proof of compatibility with another model file. Keep the exact maintainer and revision in the production record. Do not copy a Wan Animate graph into this masked-edit job simply because both use a source video.
After preprocessing, reopen the resulting source and mask rather than reviewing only the originals. A resize, frame selection or crop can create a mismatch even when the input pair was aligned. Keep the original masters and preprocessing outputs separate so the cause remains traceable.
Describe the finished shot and preserve an edit baseline
A blue hardback notebook rests on the desk and moves gently to the right. Its cover remains closed and its rectangular shape stays consistent. The lamp, framed wall artwork, desk edge and fixed camera composition remain visually continuous.
This is an original unexecuted output brief for the authored exercise. It is not an instruction that has already produced a blue notebook. Save the exact prompt, selected mask and implementation settings with each returned file. When revising a failed take, change one input at a time. A new prompt, a larger mask and a different reference together make it harder to understand why the result changed.
Review the target boundary and retained scene separately
- Watch the complete output at normal speed to assess the scene beat.
- Inspect the notebook shape, cover and movement throughout the edit.
- Follow the edit boundary for halos, remnants and sudden edge changes.
- Compare the lamp, wall artwork and desk edge with the accepted source.
- Reopen the export in the destination editor and verify actual duration, dimensions and any required audio.
Use the blank VACE edit review sheet. Record the first visible target defect, boundary defect and surrounding-scene change as separate observations. A plausible replacement prop does not excuse a bending lamp. Conversely, a clean background does not establish that the target action is correct.
Return the accepted insert to the cinema sequence
Check the shot before and after the insert. The prop's color, position and movement must connect to those neighboring shots. Use the short drama workflow for the scene beat and the multi-shot guide for coverage. For full character motion and replacement, use the separate Wan Animate workflow.
Explore Clout AI Cinema for its available project and scene workflow and Image and Video for supported visual generation. This guide does not establish VACE hosting or a local checkpoint importer in Clout. For removing an unwanted element across a shot, see the broader video object removal guide.
