Weekly Commercialization Report: Dewu Launches First C-End AI Authentication Robot at WAIC; Pop Mart’s Wang Ning Visits Apple HQ to Study AIGC Industrialization; Anta Brand CEO Xu Yang Departs Amid AI-Driven Organizational Realignment

Dewu’s 'Dewu-Jian': From Authentication Tool to Trust Infrastructure

On July 25, 2024, Dewu App officially launched 'Dewu-Jian'—China’s first AI authentication robot open to end consumers—at the World Artificial Intelligence Conference (WAIC) Shanghai. Unlike generic multimodal models, 'Dewu-Jian' is built upon Dewu’s proprietary vertical large model Dewu-VLM v2.1, trained on over 12 million high-resolution images of sneakers, apparel, handbags, and other premium lifestyle goods—each manually verified. It constructs a fine-grained visual knowledge graph covering 17 categories, 426 brands, and 28,000 SKUs. Benchmarked under natural lighting via iPhone 15 Pro, it achieves 98.7% item-level recognition accuracy on iconic products including Air Jordan 1, Yeezy 350, and Gucci GG Marmont, with false positive rate <0.3% and latency <1.2s (edge-cloud hybrid inference). Critically, 'Dewu-Jian' is embedded into Dewu’s transaction flow: mandatory AI pre-screening precedes resale listing, with results routed to human authenticators—forming a three-tier trust architecture: 'AI pre-screening + expert verification + blockchain attestation'. Peak daily authentication requests reached 473,000.

• Key tech stack: Dewu-VLM v2.1 (~12B params, 83% proprietary image data), CLIP-Adapter fine-tuning module, lightweight ONNX Runtime edge engine
• Commercial validation: Week-one launch drove 21.4% uplift in secondhand resale GMV; average wait time dropped from 4.2 hours to 37 minutes
• Compliance design: Federated learning enables local feature extraction; raw images never uploaded; reports embed CNAS accreditation IDs

Pop Mart’s Apple Visit: Benchmarking AIGC Industrialization

On July 22, 2024, Pop Mart founder & CEO Wang Ning led technical, IP content, and 3D modeling leads to Apple Park for a two-day closed-door technical exchange. Core focus: Apple’s internal AIGC toolchain 'Muse' (internal codename) for scalable IP-derived content production—including Apple Vision Pro–captured point clouds → NeRF-based high-fidelity 3D asset reconstruction → Core ML–deployed Stable Diffusion XL fine-tuned models for scene-aware rendering → Final Cut Pro AI plugins for automated audio-timing sync. Apple applies this pipeline to Apple Store retail animations and Holiday Campaign assets: $830 cost per 15-second video (vs. $12,500 outsourcing avg.), iteration cycle reduced to 4.7 hours. Pop Mart disclosed its 'Molly Factory Phase II' will integrate NVIDIA Omniverse and deploy a private SDXL-1.5 cluster with LoRA fine-tuning support by Q4, targeting ≤90-minute 3D asset generation (down from 72 hours).

Anta’s CEO Departure: Organizational Calibration Under 'AI+Design'

Anta Group announced on July 26, 2024, that Brand CEO Xu Yang departed for personal reasons, with Group VP Chen Chen assuming interim duties. This follows the November 2023 launch of Anta’s 'AI+Design' central platform strategy—consolidating previously siloed AI tools across FILA, DESCENTE, and KOLON SPORT into a unified 'Anta-Design Agent' system, fine-tuned on Llama-3-70B. The system now covers trend forecasting, fabric simulation, and pattern optimization. For example: integrating WGSN trend data + Taobao real-time text streams enables 82.6% accuracy in predicting breakout color palettes 11.3 weeks ahead; Materialise–co-developed physics engines raise 3D sampling substitution rate to 67%. Xu Yang’s exit signals a strategic pivot: future brand leadership requires prompt engineering fluency and workflow re-engineering capability—not just channel/market expertise.

HOKA × Hou Minghao: Fully AI-Generated Ad as Industrial Output

HOKA announced actor Hou Minghao as global brand ambassador on July 24, 2024, debuting the fully AI-generated film 'Gravity Shift'. Co-produced by Luma AI (NeRF video generation), Runway ML (Gen-3 text-to-video), and Adobe Firefly (texture enhancement): 2,000 HD motion frames of Hou formed the base for personalized ControlNet training; composite prompts ('floating particles during slow run', 'gravity inversion', 'HOKA Bondi 9 sole deformation close-up') generated the final 12-second 4K HDR output. Total production time: 38 hours (including human review cycles); 76% budget saved vs. live-action shoot; eliminated cross-border filming compliance risks. All shoe deformation physics validated via ANSYS finite-element simulation against HOKA’s official Technical White Paper specs.