Meituan Tackles Visual Hurdles for Million Merchants via AI Poster Generation

For business marketing, an eye-catching poster often serves as the initial touchpoint for reaching consumers. Yet, millions of small and medium-sized enterprises face persistent hurdles, including limited design resources, high outsourcing costs, and the pressure for rapid delivery during time-sensitive campaigns. Traditional design workflows are often inefficient and result in homogenized outputs. While AI generation offers new possibilities, ensuring that these images are not only visually appealing but also meet rigorous commercial standards remains a significant industry challenge.
To tackle this issue, our intelligent creation team has developed a comprehensive technical framework spanning "generation, editing, and evaluation," and has open-sourced all core innovations in the MeiGen-AI repository . This system comprises three primary modules: PosterCraft, PosterOmni, and PosterReward, designed to shift commercial poster design from "manual, complex creation" to "automated, intelligent production."
Regarding foundational generation capabilities, PosterCraft abandons the traditional modular pipeline in favor of an end-to-end unified optimization framework. By leveraging large-scale text rendering optimization and multi-stage reinforcement learning, the model resolves common "errors and omissions" when AI processes Chinese text, multi-line layouts, and complex designs, achieving near-industry-leading accuracy in text rendering. This advancement means AI is no longer merely "drawing pictures" but can accurately convey product selling points.
To handle more intricate design scenarios, PosterOmni functions as an "intelligent design assistant." It consolidates six distinct design tasks—including image expansion, completion, aspect ratio adjustment, and style transfer—into a single model, eliminating the need to link multiple tools. Whether performing local refinements based on a reference draft or restructuring global layouts, the model adapts to the visual tone of various industries, such as catering and retail, while maintaining consistency of the main subject.
Quality assessment acts as the "gatekeeper" for this system. PosterReward, the first reward model specifically engineered for poster quality evaluation, establishes automated standards covering composition, color harmony, and atmosphere. It not only facilitates large-scale online "quality inspections" but also serves as a reward signal to drive the generation model's continuous evolution, creating a self-improving "generation-evaluation-optimization" closed loop.
Currently, this technical framework has been successfully deployed in real-world scenarios, including generating meal set images for Meituan Takeout, brand IP design, and platform content governance. By leveraging technology to achieve "creative equity," even merchants without professional design expertise can produce marketing materials of professional quality within minutes. This approach not only redefines the visual marketing strategies for local life services but also offers a valuable example of AI application in specialized business sectors.
Related article
Anthropic Launches Claude Mythos to Solve Erdős Conjecture With Concise Proof
On May 26, Sholto Douglas, an engineer at Anthropic, took to X to announce that their latest model, Claude Mythos, had cracked a century-old puzzle in combinatorial geometry: the Erdős unit distance conjecture. Dubbed a "clever and concise proof," th
Amazon to Acquire Decart for $7B, Securing Key AI Chip Adapter
Per the Wall Street Journal, Amazon is negotiating to purchase AI startup Decart at an estimated $7 billion valuation. If no competing bids arise, the transaction could close shortly, underscoring major tech firms’ urgency to secure computing resourc
How Locus tackles one of robotics' biggest challenges: manipulation
NeuraGrasp handles diverse items, including small containers, porous polybags, and textiles. | Source: Locus RoboticsLocus Robotics Corp. has solved numerous robotic hurdles in live operational settings. Its autonomous mobile robots (AMRs) have trave
Related Special Topic Recommendations
Comments (0)
0/500

For business marketing, an eye-catching poster often serves as the initial touchpoint for reaching consumers. Yet, millions of small and medium-sized enterprises face persistent hurdles, including limited design resources, high outsourcing costs, and the pressure for rapid delivery during time-sensitive campaigns. Traditional design workflows are often inefficient and result in homogenized outputs. While AI generation offers new possibilities, ensuring that these images are not only visually appealing but also meet rigorous commercial standards remains a significant industry challenge.
To tackle this issue, our intelligent creation team has developed a comprehensive technical framework spanning "generation, editing, and evaluation," and has open-sourced all core innovations in the
Regarding foundational generation capabilities, PosterCraft abandons the traditional modular pipeline in favor of an end-to-end unified optimization framework. By leveraging large-scale text rendering optimization and multi-stage reinforcement learning, the model resolves common "errors and omissions" when AI processes Chinese text, multi-line layouts, and complex designs, achieving near-industry-leading accuracy in text rendering. This advancement means AI is no longer merely "drawing pictures" but can accurately convey product selling points.
To handle more intricate design scenarios, PosterOmni functions as an "intelligent design assistant." It consolidates six distinct design tasks—including image expansion, completion, aspect ratio adjustment, and style transfer—into a single model, eliminating the need to link multiple tools. Whether performing local refinements based on a reference draft or restructuring global layouts, the model adapts to the visual tone of various industries, such as catering and retail, while maintaining consistency of the main subject.
Quality assessment acts as the "gatekeeper" for this system. PosterReward, the first reward model specifically engineered for poster quality evaluation, establishes automated standards covering composition, color harmony, and atmosphere. It not only facilitates large-scale online "quality inspections" but also serves as a reward signal to drive the generation model's continuous evolution, creating a self-improving "generation-evaluation-optimization" closed loop.
Currently, this technical framework has been successfully deployed in real-world scenarios, including generating meal set images for Meituan Takeout, brand IP design, and platform content governance. By leveraging technology to achieve "creative equity," even merchants without professional design expertise can produce marketing materials of professional quality within minutes. This approach not only redefines the visual marketing strategies for local life services but also offers a valuable example of AI application in specialized business sectors.
Anthropic Launches Claude Mythos to Solve Erdős Conjecture With Concise Proof
On May 26, Sholto Douglas, an engineer at Anthropic, took to X to announce that their latest model, Claude Mythos, had cracked a century-old puzzle in combinatorial geometry: the Erdős unit distance conjecture. Dubbed a "clever and concise proof," th
Amazon to Acquire Decart for $7B, Securing Key AI Chip Adapter
Per the Wall Street Journal, Amazon is negotiating to purchase AI startup Decart at an estimated $7 billion valuation. If no competing bids arise, the transaction could close shortly, underscoring major tech firms’ urgency to secure computing resourc





Home






