Supporters of Marcus Endicott’s Patreon can access weekly or monthly video consultations on this topic.
The Chinese registered company name for Soul App (Soul) is 上海任意门科技有限公司. 上海任意门科技有限公司 is a single Chinese legal entity whose name is rendered in English in two common ways because there is no universally enforced English version: “Shanghai Anyimen Technology” is a pinyin-style transliteration that approximates the sound of 任意门 (Rènyìmén), while “Shanghai AnyDoor Technology Co., Ltd.” is a meaning-based translation that renders 任意门 as “Any Door,” producing an English name that reads naturally and conveys the concept; as a result, different sources may adopt either form for branding, press coverage, or database entries even though they refer to the same company.
Soul App (Soul) is operated by Shanghai Anyimen Technology Co., Ltd. and sits in the “digital human” landscape primarily as an avatar-first social platform that treats user interaction through digital identities as the default interface, while also productizing and publishing core enabling technology for real-time synthetic characters. It positions itself as an AI-native social product and states that all users socialize via digital avatars, supported by a self-developed large model called Soul X (Soul X) and related recommendation systems, while its AI team brand Soul AI Lab has open-sourced SoulX-FlashTalk, a real-time digital human generation model described as achieving about 0.87 seconds start-up latency and 32 fps real-time throughput at a 14B model scale, aligning the company with low-latency, long-duration, audio-driven avatar streaming for live interactive use cases rather than offline rendered “talking head” content; this open release builds on earlier public technical disclosures describing the application of “visual digital human” techniques inside the Soul product experience to blend digital-human generation with social interaction flows.
Soul AI Lab is the AI research and engineering group associated with Soul App, with a visible open-source presence that publishes model code and research artifacts in public repositories rather than only product announcements. Its recent work points to two practical directions: long-form, multi-speaker speech generation for podcast-style audio (SoulX-Podcast), and real-time, audio-driven digital human/avatar streaming intended for interactive use (SoulX-FlashTalk), with the latter positioned around low-latency start-up and real-time frame-rate output for sustained live generation. Across Soul’s broader AI disclosures, the lab’s outputs align with a platform strategy that treats internally developed models as infrastructure for “AIGC + social” features such as conversational agents, AI-assisted chatting, and virtual companionship experiences within the Soul ecosystem.
Zhang Lu (张璐) is the founder of Soul App and leads its in-house research division, Soul AI Lab, with a strategic focus on applied AI and real-time digital human technology. Under her direction, the team has pursued practical deployment of large model capabilities in commercial contexts, most recently addressing a persistent industry bottleneck in real-time digital human generation: the trade-off between high rendering quality, which has historically required expensive H800 GPU clusters, and low-cost production, which typically results in degraded visual output. In March 2026, Zhang Lu led the open-source release of SoulX-FlashTalk, described as the first model in the field to achieve sub-second latency at 0.87 seconds and 32 frames per second, supported by a 14-billion-parameter digital human model capable of stable generation across extended video sequences. The release was positioned not only as a technical benchmark for real-time digital human generation but as a marker of large model deployability in real-world business applications.