ByteDance's OmniShow turns product photos into AI sales videos

If you have ever tried making an e-commerce video with mainstream AI generators, you know the frustration: the moment an avatar tries to hold a perfume bottle or a bag of supplements, the hands twist into nightmare fuel and the product morphs into generic mush.

ByteDance and researchers from the Chinese University of Hong Kong just introduced OmniShow, an end-to-end video generation model built specifically to tackle human-object interactions in marketing and creator workflows.

Squeezing characters, goods, and voice into a single pipeline

Most video generators stumble when asked to follow multiple strict constraints simultaneously. OmniShow takes a fundamentally different path by accepting text, reference pictures, audio tracks, and pose skeletons all at once, supporting complex multimodal modes like RAP2V.

In practice, you can feed the model a photo of an influencer and a still image of any product, whether it is cosmetics, protein powder, or a curling iron. Add a voice recording and a basic stick-figure pose video, and OmniShow will animate the scene into a coherent, uninterrupted shot of up to 10 seconds.

What sets the demo results apart from existing competitors is consistency. The product packaging remains visually identical to the source image without melting, while the avatar delivers remarkably natural finger articulation, facial expressions, and accurate lip-sync matched to the audio.

Under the hood, the research paper details two key architectural tricks: Unified Channel-wise Conditioning to inject image and skeleton inputs without quality loss, and Gated Local-Context Attention to nail audio-visual synchronization. To overcome sparse multi-condition training data, the team used a decoupled-then-joint training recipe and introduced a new benchmark called HOIVG-Bench.

For those eager to run it locally, the GitHub repository is live, though ByteDance notes that the source code is currently undergoing internal review ahead of an open-source release.

Why it matters

For e-commerce merchants and creators, OmniShow could dramatically compress the cost of product showcases. Instead of booking studio time and reshooting takes, brands could soon turn a simple catalog photo and a script audio into ready-to-post promotional clips in minutes.

Strategically, ByteDance is playing to its greatest commercial strength. While OpenAI and other video heavyweights race to build broad physical world simulators, ByteDance is zeroing in on the lucrative engine of short-form commerce. Solving product fidelity and hand-object contact is unglamorous, but it solves a billion-dollar pain point.

We think OmniShow takes the right pragmatic path by prioritizing precise multi-modal control over sprawling scale. That said, a 10-second shot is still too brief for complete storytelling, and stitch-together workflows remain tricky. If models like this flood social feeds with synthetic influencers hawking goods, platforms will also face an uphill battle policing authenticity and consumer trust.