Marketing ops · Production

Netflix researches Vera and VOID: controllable AI video editing for promotional assets

The problem

Creating polished promotional video assets from raw footage requires complex edits — adding visual elements, replacing backgrounds, removing objects — that demand hours of specialized manual work. Existing AI video editing tools regenerate entire clips when editing, inadvertently altering elements that should remain unchanged and failing to account for physical scene continuity when objects are removed.

First attempt

Existing video editing models exhibit two documented failure modes: regenerating the entire video when only a specific element should change (causing unintended alterations), and removing objects without correcting the resulting physically implausible scene interactions.

Workflow diagram · grounded in source
1
Artist issues text edit instruction
trigger
“Given a source video and a text editing instruction, Vera jointly generates an edit layer and an alpha matte.”
2
Vera generates edit and alpha layers
ai_action
“Given a source video and a text editing instruction, Vera jointly generates an edit layer and an alpha matte. By design, Vera supports complex tasks such as object addition and background change, while ensuring that the pixels outside th…”
3
Compose layers with source footage
output
“These layers are then seamlessly composed with the original footage to produce the final edited result.”
4
User selects object for removal
trigger
“the user clicks on an object to remove”
5
VLM scene-interaction analysis
ai_action
“A VLM-based reasoning pipeline then analyzes the scene to identify other regions that will be causally affected, e.g., objects that will fall, collide, or change trajectory.”
6
VOID first-pass inpainting
ai_action
“VOID takes the video and the quadmasks as input and generates a physically plausible counterfactual video in which the object — and its interactions — are removed.”
7
Morphing detection and second pass
validation
“If VOID detects this failure mode, it triggers a second pass that re-runs inference using flow-warped noise derived from the first pass, stabilizing the object's shape along its newly synthesized trajectory.”
Reported outcome

Vera significantly outperforms existing baselines on content preservation in both automated metrics and human evaluations by 19 creative reviewers, and VOID was selected 64.8% of the time as best reflecting realistic scene evolution — though both remain in early research stages and not yet production-ready.

Reported metrics
VOID user preference rate64.8%
Vera content preservation vs. baselinessignificantly outperform existing baselines on content preservation
Manual editing work addressedhours of specialized manual editing work
Reported stack
VeraVOIDVLMCogVideoX-Fun-V1.5–5b-InPKubric
Source
https://netflixtechblog.com/toward-more-controllable-ai-video-editing-an-early-research-exploration-at-netflix-eb8160ed60a2
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Vera significantly outperforms existing baselines on content preservation in both automated metrics and human evaluations by 19 creative reviewers, and VOID was selected 64.8% of the time as best reflecting realistic…

What tools did this team use?

Vera, VOID, VLM, CogVideoX-Fun-V1.5–5b-InP, Kubric.

What results were reported?

VOID user preference rate: 64.8%; Vera content preservation vs. baselines: significantly outperform existing baselines on content preservation; Manual editing work addressed: hours of specialized manual editing work (source-reported, not independently verified).

What failed first in this deployment?

Existing video editing models exhibit two documented failure modes: regenerating the entire video when only a specific element should change (causing unintended alterations), and removing objects without correcting th…

How is this marketing ops AI workflow structured?

Artist issues text edit instruction → Vera generates edit and alpha layers → Compose layers with source footage → User selects object for removal → VLM scene-interaction analysis → VOID first-pass inpainting → Morphing detection and second pass.

Related marketing ops cases