Two former ByteDance executives have launched well-funded AI video generation startups, but independent testing shows their self-developed models still lag significantly behind ByteDance's Seedance technology, raising questions about whether superior funding alone can overcome entrenched model advantages. PixVerse Tech's Paiwo AI and Eanyo Tech's LibTV both launched in 2025 and 2026 respectively, attracting billions in venture capital from investors including Alibaba, Tencent, and Ant Group. Yet when tested side-by-side generating the same video prompts, the products revealed a fundamental challenge: building a competitive large language model from scratch is harder than building a company around an existing one.

Why Are These Startups Struggling Despite Massive Funding?

Wang Changhu, founder of PixVerse Tech and former head of AI Lab Vision at ByteDance, built the visual technology system for Douyin and TikTok from the ground up. His new company completed Series C financing in July 2026, bringing total funding to 2.98 billion RMB. Chen Mian, founder of Eanyo Tech and former global commercialization lead for CapCut, raised nearly $300 million in Series B+ financing, giving his company a valuation exceeding $2 billion.

Despite this firepower, the two companies took fundamentally different technical paths. Paiwo AI relies on a self-developed model called Pixverse V6, while LibTV functions as an aggregation platform that uses both ByteDance's Seedance and Kuaishou's Kling models alongside its own tools. This strategic difference reveals the core problem: building a competitive proprietary model requires not just money, but years of iterative research and massive computational resources.

When Tech Planet tested both products with identical prompts, the differences became stark. For a request to generate an anime-style video of Elon Musk watching a FIFA World Cup match, Paiwo AI's output featured an incorrectly proportioned character with a jawline that bore little resemblance to Musk, and the background looked more like an esports arena than a stadium. When asked to add a goal-scoring scene, the character's position suddenly shifted onto the football pitch itself, creating a logically inconsistent sequence.

LibTV, by contrast, generated a character that closely matched Musk's appearance, wearing appropriate attire for the scene, with facial expressions aligned to the plot. The model maintained logical consistency when asked to add additional elements. In a second test requesting a robot dancing ballet on stage, Paiwo AI failed to clearly depict the stage element, while LibTV fully rendered the setting.

What Makes ByteDance's Model So Hard to Replicate?

The technical gap points to a deeper advantage ByteDance possesses: not just the model itself, but the infrastructure and data advantage that comes from operating TikTok, Douyin, and CapCut. Seedance 2.5, which launched its public API on July 16, 2026, represents a significant leap forward. The model can generate continuous, unstitched 30-second video clips in a single pass, something no other commercially available model could do before that date.

This capability relies on a Sparse Diffusion Transformer architecture that extends the model's working memory across the full clip duration. Standard video diffusion models process frames using attention windows limited to short temporal spans, which causes character drift and lighting inconsistency over time. Seedance 2.5 uses sparse attention to sample across the full temporal extent of a clip, maintaining coherent scene state across every frame in a single inference pass.

The model also accepts up to 50 multimodal reference materials, including images, video clips, and audio, in a single generation request. This dwarfs prior generations: Seedance 2.0 accepted only 15 reference assets combined. For multi-character brand content or narrative sequences with full casts, the difference between 15 and 50 reference slots determines whether all key visual anchors can be held simultaneously or whether production teams must make tradeoffs.

How Are These Companies Trying to Compete?

The two startups have adopted different strategies based on their founders' backgrounds. Wang Changhu's technical expertise led Paiwo AI to focus on model research and development, betting that superior underlying technology could eventually match ByteDance's capabilities. Paiwo AI targets a broader user base with simple operation and no technical barriers, focusing on vertical-screen creative content suited to short-form video platforms.

Chen Mian's commercialization background shaped LibTV into a professional-grade platform with multi-model integration and sophisticated workflow design. LibTV's built-in skills focus on long-form video styles with higher production ceilings, catering to professional users who need flexibility across multiple models rather than reliance on a single proprietary engine.

The financial performance data reveals which strategy is gaining traction. PixVerse Tech's Paiwo AI and its overseas version Pixverse have attracted over 150 million global users across 177 countries, but monthly active users have remained around 15 million with annualized revenue exceeding $40 million. Eanyo Tech's portfolio of products shows stronger commercial momentum: LiblibAI accumulated over 30 million users with annualized revenue exceeding $300 million, while Lovart achieved $80 million in annualized revenue only five months after launch. LibTV gained over 100,000 users on its first launch day, with monthly revenue increasing more than 13 times within two months.

How to Evaluate AI Video Tools for Your Needs

• Model Architecture: Determine whether you need a single proprietary model optimized for specific use cases or a multi-model platform that offers flexibility across different generation styles and technical approaches.
• Reference Asset Capacity: Consider how many visual anchors you need to maintain consistency across your content; models supporting 50 reference materials offer significantly more control than those limited to 15.
• Workflow Complexity: Assess whether your team requires simple, user-friendly interfaces for quick content generation or professional-grade workflow tools with node-based editing and skill combinations.
• Video Duration Requirements: Evaluate whether single-pass 30-second generation meets your needs or whether you require longer clips that may need stitching and post-production compositing.
• Commercial Roadmap: Review pricing structures and intended use cases; some tools target short-form social content while others focus on e-commerce advertising, brand videos, and professional marketing content.

The broader story here is not about which startup will win, but what the competition reveals about AI development economics. Building a world-class video generation model requires not just capital, but years of research, access to massive training datasets, and computational infrastructure that only a handful of companies can afford. ByteDance's advantages stem from operating multiple platforms that generate billions of hours of video content annually, providing training data and real-world feedback loops that startups cannot easily replicate.

Meanwhile, Alibaba is consolidating its own multimodal AI capabilities under a unified command structure. In July 2026, Alibaba merged its Tongyi Wanxiang team into the Future Life Lab led by Vice President Zhang Di, bringing three multimodal models under single oversight: Wanxiang for text-to-image-and-video, HappyHorse for video generation, and HappyOyster for world models. Zhang Di, who previously led Kuaishou's Kling project before joining Alibaba, is positioning the company's AI capabilities directly into e-commerce scenarios, where video generation, image generation, and world models serve as content productivity tools for advertising, product display, and immersive shopping experiences.

The competitive landscape is consolidating around two dynamics: companies with massive platform data advantages (ByteDance, Alibaba, Kuaishou) are building proprietary models, while startups are either building specialized aggregation platforms around existing models or attempting the much harder task of competing on model quality itself. The funding these startups have raised is substantial, but the technical gap they face suggests that venture capital alone cannot overcome the structural advantages of companies with billions of hours of training data and years of iterative research already complete.