What's Next: Real-Time Video Avatars in AI Girlfriend Apps

What a real-time video avatar would actually mean

I want to be precise about the feature I'm describing, because "AI video" already means different things across the 129 platforms I track. Most AI video generation today produces a short clip: you request it, wait a bit, and get back a pre-rendered video of your character. A real-time video avatar is a different, harder thing entirely: a live, continuously rendered face that moves, reacts, and speaks back to you during an ongoing conversation, closer to a video call than a video you requested and received.

That distinction matters because it changes the entire technical problem. Pre-rendered video generation is a "generate this, then deliver it" pipeline. A real-time avatar has to render, animate, and speak fast enough that a delay feels like a video call, not a slideshow, which is a fundamentally different latency and cost problem.

Where video generation actually stands today

Right now, 22% of the 129 platforms I test offer AI video generation in some form, which is real, meaningful adoption for a feature category that barely existed a couple of years ago. But nearly all of that adoption is the pre-rendered clip model, not a live avatar. I've broken down exactly which platforms offer video and how it's implemented in a separate, more detailed count, and the short version is that video generation today is still fundamentally a "wait for a result" feature, not a "have a live conversation" feature.

What would need to be true for real-time avatars to actually work well

Three things would need to come together at once, and right now none of them are close to solved industry-wide. First, voice needs to be fast and natural, since a real-time avatar without good voice is just an animated face saying nothing useful; today voice averages just 1.81 out of 5 across the industry, the weakest category I score. Second, video rendering needs to happen with low enough latency that it feels like a live conversation rather than a buffering video call, which is a much stricter constraint than the "generate a clip and deliver it in a minute or two" model most video generation uses today. Third, the whole pipeline (understanding what you said, generating a response, animating a face, and speaking it back) has to run cheaply enough per minute that a company can actually offer it at a price people will pay.

That third point is easy to underrate. Real-time video generation is computationally expensive compared to text or even a single generated image, and running it continuously during a live conversation, rather than once on request, multiplies that cost significantly. I'd expect the first real-time avatar features to launch as a premium, metered add-on rather than something included by default, precisely because of that cost structure. Until that arrives, our best AI girlfriend ranking already tracks which platforms lead on the voice and video features avatars would eventually build on.

What real-time avatars would actually unlock, if built well

If a platform genuinely nailed this, I think it would be the single biggest jump in perceived "realness" this category has seen since large language models replaced scripted chat in the first place. Text conversation, even good text conversation, still requires you to imagine the other side of the exchange. A live, responsive face closes a meaningful part of that imagination gap, which is exactly why I think it's the most natural next frontier once voice and video generation both mature past where they currently sit.

It would also raise the bar on everything underneath it. A mediocre real-time avatar, one with obvious lag, robotic voice, or a face that doesn't quite match what's being said, would probably feel worse than no avatar at all, similar to how a bad voice feature can feel worse than sticking to text. This is a feature where getting it half right is genuinely worse than not shipping it yet.

How I think this would actually get priced

Given how expensive continuous, real-time rendering is likely to be compared to a single on-demand video clip, I'd expect real-time avatar features to launch as a metered, premium add-on rather than something folded into a standard subscription from day one. That mirrors how video generation itself launched: a premium feature on top of an existing chat subscription, not a replacement for it. Only 2 of the 129 platforms I currently track price as premium overall, so I'd expect this specific feature to show up as a paid add-on inside an otherwise budget-to-mid-range platform, rather than justifying an entirely new premium tier of platform by itself.

I'd also expect early pricing to be usage-based, minutes of live avatar time, rather than a flat monthly add-on, simply because the compute cost scales directly with how long you're actually using it, unlike a chat subscription where marginal cost per message is much smaller and easier to average into a flat price.

I'd expect the first implementations to cap usage tightly, a handful of minutes per day or per week, rather than offering unlimited real-time avatar access even to paying subscribers, purely because the cost curve doesn't support unlimited use at a price most people would pay. That's a reasonable, honest way to launch a genuinely expensive feature, even if it means the earliest version feels more like a rationed preview than the always-on presence fiction tends to depict.

How this is different from the broader video-generation shift happening industry-wide

It's worth separating this specific feature from the bigger, industry-level story of generative video reshaping the companion app business more broadly. Real-time avatars are a narrow, specific feature: a live face during a single conversation. The broader industry shift, more platforms adopting some form of video generation at all, new business models built around it, is a wider trend that I cover in a separate piece on how generative AI video is reshaping the companion app industry. Real-time avatars are one specific, harder-to-reach destination inside that larger trend, not the whole story of it.

If you want the technical explainer on how the current, pre-rendered version of AI video generation actually works today, and why it only recently became technically feasible at all, I've written a separate deep dive on how AI girlfriend video generation works and why it's so new, which is a good foundation for understanding why the real-time version is still a step further out.

My honest, informal timeline for this

I don't think real-time video avatars are a "next year" feature for most of the 129 platforms I track, but I also don't think they're a decade away. My honest opinion is that this sits behind voice quality catching up first, since a company solving live avatar rendering without also solving natural, low-latency voice would be shipping a feature that undercuts itself. Once voice interaction meaningfully improves past its current 1.81 out of 5 average, I'd expect real-time avatars to follow within a couple of development cycles from whichever platforms are already furthest ahead on video generation and voice combined.

I go into the wider industry trajectory, including how fast this category has moved historically, in my broader look at where AI girlfriends are actually headed next. You can also read more about how I test and score every platform or my background as a researcher covering this space.

Further reading

• Video Generation Adoption Across AI Girlfriend Apps: A Platform-by-Platform Count
• How AI Girlfriend Video Generation Works (And Why It's So New)
• The best AI girlfriend apps, ranked