CoInteract generates influencer-style product demos from two photos

Shooting a quick influencer-style product demo used to require samples, lighting, a creator, and multiple takes. While generative AI promised to replace that hassle, current video models almost always trip over simple physical interactions: hands morph unnaturally, and fingers clip right through the product like a bad video game glitch.

Researchers introduced CoInteract (also referred to as Co-InAct in early previews), an end-to-end framework designed specifically to fix these awkward physical failures. You feed it just two still images—one photo of a person and one clean shot of a product—along with a text prompt or spoken audio, and it outputs a realistic video of that person naturally presenting and handling the item.

Teaching AI the rules of physical contact

Current diffusion models excel at photorealistic textures, but they consistently struggle with structural stability in hands and faces, as well as plausible hand-object boundaries. CoInteract tackles this directly inside a Diffusion Transformer (DiT) backbone using two novel mechanisms.

First, it uses a Human-Aware Mixture-of-Experts (MoE) setup. Through spatially supervised routing, the model sends tokens to lightweight expert networks dedicated to delicate regions like fingers and facial features, preserving structural fidelity without ballooning parameter size.

Second, the researchers developed Spatially-Structured Co-Generation, a dual-stream training framework. One stream handles the RGB appearance, while an auxiliary stream models the physical interaction geometry between the person and the object.

During training, this structural stream cross-references RGB tokens and forces the shared backbone to learn proper collision boundaries. When it is time to actually generate a video, the structural branch is simply dropped, yielding zero-overhead inference with realistic physical contact.

Users can also feed step-by-step instructions in their prompts, prompting the virtual presenter to perform sequential product demonstrations smoothly.

Why it matters

For digital marketers and e-commerce merchants, CoInteract points to a future where producing UGC-style promotional content requires virtually zero production budget. Instead of shipping samples and waiting weeks for creators to deliver footage, brands could generate dozens of tailored product showcase variants within minutes from static product listings.

In our view, this development is a serious warning shot for lower-tier UGC creators and micro-influencers. If all a merchant needs is a licensed face photo and a packshot to churn out convincing demos, basic influencer gigs will dry up fast.

Technically, CoInteract makes a very clever trade-off. Rather than hoping a massive generalist model eventually deduces intuitive physics on its own, it forces physical geometry priors during training and strips them at runtime to keep generation fast. While it is currently an academic paper without immediate one-click public access, commercial video platforms will almost certainly adopt this dual-stream approach soon.