TL;DR

• Ad insertion has always been a hard cut. I wanted to explore whether generative AI could change that.
• At Bitmovin’s Q2 2026 hackathon with ByteDance, I built two approaches: AI product placement inside scenes, and seamless AI-generated transitions between content and ads.
• ByteDance made their BytePlus model suite available to all teams, which made the whole thing possible in a hackathon timeframe.
• The models are not ready for full automation yet, but the pipeline works and the results are worth paying attention to.

Advertising in video has not changed much in decades. You watch content, it stops, an ad plays, and then your content resumes. For viewers, that interruption breaks immersion and trains people to tune out the moment a cut happens. For brands, it is arguably just as bad: an ad that lands in the middle of a jarring transition is not one that gets remembered for the right reasons. The assumption baked into this model is that interruption is just the price of doing business on both sides. I think generative AI gives us a real opportunity to question that.

In this post I walk through the two approaches I explored at the hackathon, how I built them, where they worked, and where the models still have room to improve.

Approach One: Regenerating Scenes with Injected Products

The first approach was to take an existing video, identify moments where a product could logically appear, and use generative AI to rebuild those scenes with the product naturally embedded in the frame. The workflow relied on three BytePlus models from ByteDance, each handling a distinct part of the pipeline:

• Seed: visual scene analysis, understanding lighting, surface materials, and spatial context within each shot
• Seedream: generating preview images showing how a product would look placed within a scene
• Seedance: the actual video generation and editing, rebuilding the scene with the product in place

Using scene boundary data from Bitmovin’s AI Scene Analysis, I split a video into scenes and passed each through Seed’s visual understanding model. Seed analyzed the environment, identified candidate surfaces or positions where a product could plausibly sit, and rated each opportunity for contextual fit based on lighting, scene category, and spatial position. Those opportunities were then surfaced through a review UI showing the scene context, proposed product, target position, lighting conditions, and a confidence score. From there I could approve, reject, or edit the Seedance generation prompt directly before kicking off a render.

The product placement review interface, showing scene context, lighting analysis, confidence score, and approve/reject controls.

I tested this with a mountain trekking scene. A Red Bull can was matched at 85% confidence to a shot of a lone hiker traversing a snowy alpine landscape, with the system reasoning that an energy drink is contextually appropriate for a high-altitude trekker. A second opportunity placed Ray-Ban sunglasses on the hiker’s face at 90% confidence, based on the highly reflective snow environment. I rejected that one in the demo, but the model’s reasoning was genuinely sound.

An 85% confidence match: Red Bull placed in the hiker’s hand during a desolate mountain sequence.

A rejected 90% confidence placement: Ray-Ban sunglasses on the hiker. High score, but I passed on this one.

Where It Hit Its Limits

Seedance was a mixed performer here. Both its editing and generation modes produced results that fell short of what I needed, with inconsistent character motion and camera tracking in the regenerated clips. I tested Luma as an alternative, and it handled motion tracking significantly better. In one test clip the camera movement and character motion held together well, though the product itself was not rendered convincingly. The broader lesson: the models are improving fast, but full automation is not there yet. Keeping a human in the loop at every placement decision, with the ability to tune the generation prompt before anything renders, is the right call for now.

Approach Two: Seamless Ad Transitions

Rather than rebuilding entire scenes, my second approach tackled a narrower problem: making the cut between content and an ad feel intentional rather than jarring. The idea was to generate a short bridging video, a few seconds of footage that picks up the visual and motion language of the original scene and carries it smoothly into the ad.

The core concept: the last frames of the original video and the first frames of the ad, bridged by a generated transition segment.

My first iteration exposed a specific problem: camera movement. When the original scene has a particular motion, such as a slow dolly forward or a pan tracking a subject, a generated transition that ignores that creates a visible jolt even if the visuals blend. I added a camera movement analysis step to fix this. Just like was done for the first few seconds of the ad, the system passes the closing seconds of the original clip through Seed, which describes the camera behavior in detail, including direction of movement, speed, focal tracking, and depth of field. That description then conditions the Seedance prompt so the transition matches the motion dynamics of the scene it is bridging from.

The refined workflow came down to three steps:

• Select a scene switch point and an ad to insert
• Extract frames from the transition window and analyze camera movement using Seed
• Generate the transition video with Seedance conditioned on the camera motion description, then stitch everything together

The final demo clip, tagged “Never Stop Exploring,” was the output I was satisfied enough to present. The motion carried through, the transition felt motivated rather than mechanical, and the ad appeared without the usual hard break in visual continuity.

Demo of how the ad transition works within the hackathon project

What This Points To

These two approaches represent different near-term bets on where AI-assisted advertising in video goes next.

Scene regeneration has the higher creative ceiling. If generative video quality continues its current trajectory, placing contextually matched products directly into content could become technically viable within a short window. The harder challenge is infrastructure, building a pipeline that handles scene analysis, product matching, reviewer workflows, and generation at scale, and that sits squarely in the domain of video platform technology.

Seamless transitions are more practical to deploy right now. They only require generating a few seconds of bridging footage rather than modifying content, and the quality bar is meaningfully lower. A transition that feels like part of the content rather than a break from it changes the viewing experience in a way that is hard to achieve through targeting or scheduling alone. The ByteDance partnership and access to the BytePlus model suite made it possible to explore both directions in a single hackathon, and the results show the tooling is already worth taking seriously.