JD.com 618: Have Digital Humans Evolved Again?

A Show With No Real People, and a Systemic Upgrade for Digital Livestreaming

👦🏻 Author: Yitao

🥷 Editor: Koji

🧑‍🎨 Layout: NCon

On June 1, Children's Day, JD.com put on a show with no human performers — the "JD 618 Cyber Gala," billed as China's first all-digital-human AI shopping livestream event.

The gala brought together classic IPs from multiple generations. The Monkey King, Afanti, and Big-Headed Son shared the stage in the musical Scroll of Time, cramming childhood memories from several generations into a single performance.

A capybara did stand-up comedy. JD Gaming esports player Paraboy appeared in digital form to chat across dimensional boundaries. The Haier Brothers and Joy sang together, while a JD.com delivery worker digital human performed a street dance routine called Express Delivery.

The finale was the classic group sing-along, with multiple digital humans taking the stage together to perform "AI Together."

This purely digital-human production, with no human actors involved, delivered a commercial performance far exceeding market expectations. The gala drew over 4.9 million viewers that night and successfully drove a 300% year-on-year increase in brand sales.

The technology behind it came from JD.com's digital human platform, JoyStreamer.

Take the Haier Brothers' song-and-dance number Rhapsody on the Green Field. They weren't just lip-syncing the Leo song — they performed various soccer moves in sync with the plot and musical rhythm, with the characters' tone, expressions, and body movements forming coherent, unified responses without any jarring disconnect.

Or consider the cyber fashion show segment. The digital humans' capabilities went far beyond basic lip-syncing. From the models' confident, fluid runway walks and the tension of their pivot turns, to the delicate refraction of fabric textures under lighting and the realistic dynamics of hemlines swaying with each step, the visual presentation reached a remarkably high standard, making the entire show deeply immersive.

Take the earth-sign group as an example. The setting was a Baroque waterscape runway, with a white catwalk winding across the water's surface, as digital models in earth-toned suits and trench coats walked slowly toward the camera.

The visuals speak for themselves — the immediate reaction is: this looks good.

Roll back two years, and it wasn't just JD.com — the entire industry's digital livestreaming was stuck at an awkward starting point. Hosts could only sit in front of the camera reading scripts, turning their heads with mechanical stiffness, let alone performing movements, changing scenes, or coordinating with product displays.

Back then, digital humans solved the problem of "whether someone is broadcasting." Whether they broadcasted well or convincingly was another matter entirely. From "sitting and reading scripts" to runway-level performance — in two years, digital humans crossed a significant technical divide.

But however impressive a gala may be, it's ultimately a concentrated showcase. The fashion show proved JD.com's digital humans could handle complex scenarios with compelling presence. The real place where digital human value gets tested, however, is in the tens of thousands of ordinary merchant livestream rooms.

If the past decade of e-commerce competition unfolded around efficiency optimization of "people, goods, and venue," then this wave of digital human upgrades represented by JD.com's digital humans is transforming "people, goods, and venue" themselves into systematically generatable and schedulable digital assets.

One Host, One Team

In the livestream e-commerce business, there's an invisible dividing line: whether you can run a livestream room well often depends less on how much traffic you buy than on whether you've assembled a reliable team.

Traditional digital livestreaming primarily solved the "presence problem" for broadcast scenarios: filling gaps when merchants lacked hosts, reducing labor costs, preventing interruptions and dead air — essentially achieving basic broadcasting functionality.

But livestream competition has evolved to a point where merchants need far more than "someone broadcasting." They need to explain products clearly, retain users, manage interactions, and ultimately sustain operational efficiency improvements.

This 618, JD.com's digital human JoyStreamer underwent a systemic upgrade around the core conversion chain of e-commerce livestreaming. From a digital host that "can broadcast and answer," it became a complete AI livestream team.

Pre-Broadcast: Hand the Prep Work to AI

Before going live, the system performs pre-broadcast diagnostics across dimensions including product selection, set design, and copywriting, alerting merchants to areas needing optimization.

Scripts don't need human writers either. The intelligent scriptwriting feature automatically generates spoken broadcast scripts based on product information.

The COFCO Group case is fairly typical. During the Western China International Fair in May this year, JD.com digital humans built 24/7 livestream rooms for 13 of COFCO's self-operated stores. From Andalucía to Malin, each store had different products, and the system generated a unique script set for each one.

The barrier to creating a digital human host is also extremely low. Merchants simply upload a front-facing portrait photo or a clip of real-person livestream footage, and the system quickly generates a "digital twin" highly consistent with the real host.

Anyone who's done e-commerce livestreaming knows that a broadcast's success has considerable randomness. If the host is in good form that day, they might really resonate with users. But human condition fluctuates — yesterday's best performance may not be repeatable today, and viewers won't wait for you to get your act together.

Now, a successful broadcast can be "fixed" — becoming a digital asset that's callable at any time and can be refined repeatedly. Livestream operations are shifting from "gone after broadcasting" toward "sustainable accumulation."

During Broadcast: Orchestration and Floor Control Are the Core of This Upgrade

First, orchestration:

Previously, digital humans in livestream rooms basically read from scripts with fixed camera angles and monotonous pacing, moving from one product to the next.

But in reality, a top-tier influencer livestream typically has an entire team behind it: a director switching camera angles, an assistant monitoring bullet comments, an operator managing rhythm, and a floor controller adjusting atmosphere.

After the upgrade, what JD.com digital human JoyStreamer does is essentially compress an entire team's work into one system: unified scheduling of digital human movements, product display footage, special sound effects, co-host dialogue, and marketing visual content — sound, form, and meaning all integrated together.

A recording from a beauty category livestream room illustrates what this orchestration looks like in practice.

When the host reaches a core selling point, the camera simultaneously cuts to a product close-up.

The system can even automatically generate background KT board visual cards displaying key parameters, dynamically switching content based on the currently featured product.

When promotional moments arrive, red envelope rain effects and bullet comment effects trigger on schedule, fully replicating the atmosphere of a real-person livestream room.

Watching this segment, if no one told you in advance this was AI broadcasting, you might not immediately realize it.

Previously, a small or medium merchant wanting to build a livestream room at this level would need at least a 3-5 person team, plus considerable time to gel.

The issue isn't just cost. The deeper barrier is experience. The orchestration capability of top livestream rooms comes from directors, operators, and floor controllers grinding through hundreds or thousands of broadcasts — experience and tacit understanding that can't be solved by temporarily hiring people with money.

Imagine a husband-and-wife beauty brand. The wife herself is their best host. She understands the products, understands the users, has charisma. But she can only broadcast four or five hours a day, because after going offline she still needs to pack orders, reply to messages, and coordinate with suppliers. She can't afford a professional livestream team, let alone maintain a floor controller who understands conversion tactics. Her livestream capability is trapped within the hours she can stay alert.

Now, a system runs and that's it. Merchants don't need to assemble their own teams to achieve orchestration standards approaching top-tier livestream rooms.

Now, floor control:

The key change is here: digital humans no longer wait to be asked questions — they start proactively managing the livestream room.

For example, a user enters the livestream room, asks nothing, just watches. Previously, the digital human wouldn't engage because without keyword triggers, there was no response.

Now the system judges that this user may be hesitating, proactively guiding with a comment or directly cutting to the core selling points of the featured product.

When several questions flood in simultaneously via bullet comments, the system independently judges which are high-conversion questions and prioritizes answering those.

When traffic suddenly spikes, it can adjust pacing. And not just dialogue — camera switching, product display order, and conversion tactics all coordinate in sync. The entire livestream room upgrades from fixed-script playback to an intelligent broadcast environment that responds in real time.

This global scheduling capability takes real human floor controllers extensive time in livestream rooms to develop, and however experienced they are, they can't simultaneously monitor hundreds of bullet comments and make real-time decisions.

This floor control capability, viewed industry-wide, is the least replicable part of JD.com's digital humans.

AI livestream tools that do proactive floor control aren't unheard of on the market. But JD.com's system scheduling logic rests on two unique foundations.

First, JD.com's depth of understanding of user purchase decisions. JD.com's self-operated model means the platform doesn't just know what users "browsed and clicked" — it has the complete transaction chain from browsing, adding to cart, placing orders, through delivery, returns, and repurchases, enabling complete insight.

For instance, a user who has purchased a certain brand's facial cleanser three times on JD.com, started price-comparing a week before each major promotion, with average order values concentrated in the 150-200 yuan range. This structured consumption history means when the system judges "this user may be hesitating," it's basing this not on traffic tags but on actual decision-making habits.

Second, the injection of JD.com's procurement and sales system experience. JD.com's procurement and sales teams manage both product selection and pricing and promotional rhythm, with a feel for livestream sales pacing accumulated through tens of millions of real transactions. This isn't merely large model capability — it's the datafied expression of JD.com's decade-plus of e-commerce operational experience.

Based on these two layers of accumulation, JD.com digital humans are transforming key capabilities of top-tier influencer livestreaming into AI capabilities that merchants can access at low cost, replicate at scale, and continuously optimize.

One detail deserves separate mention.

The digital human pauses mid-sentence to take a sip of water. This action has no functional purpose, but it's a microcosm of the digital human's "free state" generation.

Supporting ultra-long uninterrupted broadcasts with consistent identity, plus these seemingly functionless but deeply humanizing small gestures, constitutes the natural experience that makes viewers feel "this is a real person broadcasting."

Behind these capabilities is an architectural-level change. JD.com digital human JoyStreamer has upgraded from a single AI host to a multi-AI Agent collaborative system, with different Agents each responsible for one segment: some manage script rhythm and product explanation, some monitor bullet comment interactions and user behavior, some handle floor control strategy, promotional timing, and conversion moments.

What you see on screen is one host, but behind the scenes is a team of Agents working in coordination.

Post-Broadcast: The Final Link in the Operations Loop

After the livestream ends, the system automatically generates operational reports, analyzing performance against set goals and providing strategic recommendations.

For many small and medium merchants, "post-mortem review" is the most easily skipped step. Often after finishing a livestream they're already exhausted, with no manpower left to review data and think strategy.

With this process automated, each review's conclusions feed back into the next broadcast's diagnostic phase. From pre-broadcast diagnosis through post-broadcast review, the chain completes, forming a continuously optimizing cycle.

This entire capability set points toward one goal: giving ordinary merchant livestream rooms a "top-tier influencer-level" experience.

The Capabilities of the Few, Tools for the Many

The evolution path of JD.com's digital humans encapsulates the industry's transformation.

• Early on, it was idle-time coverage. During periods when merchants lacked hosts, digital humans filled in, solving the "whether someone is broadcasting" problem.

• Later came founder IP digital humans, brand president digital humans, with digital humans beginning to carry brand image and traffic gateway roles.

• Still later, commercially viable digital humans, with expanding service scale, as more merchants treated digital humans as normalized operational tools.

This year, it's AI-native digital humans. With improvements in multi-Agent coordination, proactive floor control, and intelligent diagnostics, the digital human role shifts from "filling hours" to "managing operations."

This year's 618 opening data confirms this: JD.com digital human merchant service scale exceeded 80,000, cumulative broadcast time surpassed 800,000 hours, and overall GMV exceeded 240 million yuan.

Behind these numbers is an industry shift worth seeing clearly.

In recent years, e-commerce livestreaming has had a structural contradiction: livestreaming has proven to be one of the highest-conversion scenarios, but it's also the most human-dependent.

A top-tier influencer livestream can drive GMV in the tens of millions, but this capability is tightly bound to the individual, difficult to replicate, and impossible to scale. Large numbers of small and medium merchants know livestreaming matters but are stuck at the threshold of "no good host, can't afford a team."

What JD.com digital human JoyStreamer's upgrade touches is the core of this contradiction: for the first time, it transforms the implicit know-how of top-tier influencer livestreaming into explicit, callable system capabilities. The orchestration rhythm, floor control judgment, and user insight that previously belonged only to a handful of experts are beginning to be沉淀 into the platform.

For merchants, the barrier to livestreaming capability is dropping. Large brands can cost-effectively deploy multiple livestream rooms without staffing professional teams for each one. Small and medium merchants get their first opportunity to use operational capabilities that previously belonged only to the top tier.

For users, the viewing experience across livestream rooms will trend toward consistency. In any livestream room, they can get clear product explanations, timely interactive responses, and reasonable promotional pacing — no longer needing to rely on luck when entering a room.

For the industry as a whole, when the "broadcasting" link of livestreaming is leveled by AI, competition returns to more fundamental grounds: whose products are better, whose supply chain is stronger, whose service is more solid.

JD.com previously won the 2024 Wu Wenjun Artificial Intelligence Science and Technology Award Grand Prize for "Multi-modal Interactive Digital Human Key Technologies and Industrial Applications." This is China's highest technical honor in AI, and the only grand prize that year. From foundational technology to industrial application, this award represents concentrated recognition of the complete technical accumulation behind JoyStreamer.

Looking further ahead, digital human livestreaming is completing an identity transformation: from early-stage "cost-reduction tool" to "operational infrastructure" for e-commerce. Orchestration, floor control, diagnostics — these implicit capabilities that previously existed only in top teams are being abstracted into standardized AI modules with scalable replication potential.

The Cyber Gala was an extreme stress test. Ninety minutes uninterrupted, multiple IPs sharing the stage — demonstrating that this capability's ceiling is sufficiently high.

Returning to daily livestream rooms, more categories and longer durations still need verification. But the direction is already clear: top-tier influencer experience is being systematized into infrastructure. When these capabilities are no longer exclusive to the few, more merchants' opportunities will open up.

Crossing is seeking independent contributors to write AI product and model reviews.

If you've written articles like: "Hands-on with PixVerse C1", "Hands-on with LibTV", please contact zeo0811@gmail.com. Your email should include: ① personal introduction, ② AI review articles you've written.

We offer competitive compensation. Looking forward to observing and documenting the AI era together with you 🎪