iQIYI’s 580-Day AI Push: From ‘Skip Watch’ to Script and Image Workshops

In the demonstration, a scene in which a painter leaves an unfinished work in a temple and a black-clad intruder dies in a thunderclap was reduced from eight minutes to three, keeping the suspense while making the opening play like a short drama. Gong proposed Skip Watch three months earlier, based on his view that different users focus on different parts of the same content and that attention is becoming scarcer. iQIYI’s approach uses a large language model to analyze individual viewing behavior and identify and recommend plot points, allowing viewers to move between a long drama and a shorter version.

About 1,400 kilometers from Beijing, in Hengdian, director Jin Zheyong was shooting a scene in which the actor appeared to cross mountains, grasslands and deserts. The actor was running on a treadmill inside a studio in front of a 500-square-meter immersive LED wall, with no green screen and no elaborate set. All the background footage was generated through AI text-to-image and image-to-video tools; from the director’s approval of the first concept image to the completion of video backgrounds for three scenes took less than a day.

The article frames these changes as part of the 580 days since Gong said generative AI would completely change the film and television industry in three to five years. At the Harvard Business Review China 2023 annual meeting, Gong concluded his speech with the line: “Humans will not be replaced by AI. Humans who do not master AI will be replaced by humans who do.”

Inside iQIYI, the shift began in March 2023, shortly after GPT-4 was released, when Zhang Xiaomei, head of the company’s film and television literature research and development center, was pulled into a high-level meeting with Gong and Liu. The discussion centered on using AI to reshape film and television content creation and evaluation. Gong told Liu that their goal was to “replace” Zhang, she recalled. In her new role, Zhang became a teacher for the AI model, working to teach it what constitutes a plot and a narrative structure. Every Tuesday after work, her team met with product development and algorithm teams to read scripts and discuss and rate plot points. Within two to three months, they rated scripts including “The Bad Kids” and “Love Between Fairy and Devil.” The algorithm team used the material to convert the scripts into structured information that the model could read, training it to move from literal recognition to understanding context and complex emotions. The resulting application, “Script Workshop,” was released in 2024.

The push extended across the company. On March 3, 2023, all employees were required to attend a training session titled “Playing with AI in iQIYI business scenarios,” the only all-staff training in iQIYI’s 15-year history. The company opened an AIGC section in its internal OA system with free generative AI tools, and Gong encouraged employees to use them. Human resources ran more than 40 online courses, with more than 90% of instructors from inside iQIYI. In 2024, an iQIYI AI visual creative contest drew more than 200 cross-department teams and 121 generative AI-assisted designs; the first-prize team said it generated thousands of AI images, built more than 10 ComfyUI workflows and tried nearly 20 AI tools.

In the summer of 2023, as consumer AI applications gained users, iQIYI debated whether to follow. It decided not to build simple consumer products by calling basic large-model functions. “We judged that merely calling the basic capabilities of large models could not form iQIYI’s core competitiveness,” Liu said. “Generative AI should be used in areas that can more profoundly change the future of the film industry.” The company has said it does not develop general-purpose large language models, but its multimodal scenarios create heavy demand for model training and application. By June 2024, more than 98% of employees had used large models, and calls in 2024 were 83 times those in 2023. On its internal agent development platform, Qizhi, 64 business units used more than 6,300 agent applications.

Wang Qingfeng, who joined the newly established AIGC Content Technology Innovation Center in summer 2023, helped launch tools including Script Workshop and Image Workshop. The first version, launched on September 28, 2023, generated novel outlines and character biographies. It was iterated to extract story outlines, graded synopses and character relationship charts, with accuracy above 90%. The center also developed “Film and Television Knowledge Q&A,” which contains humanities and social science texts, and “Story Bank,” which holds 800,000 story prototypes drawn from case files for suspense and crime creators. Lu Xiangju, who worked on Story Bank, said the case files were often 4,000 to 5,000 characters long and difficult to read. The tool uses models to extract case developments, character traits, key evidence and methods into readable stories. “Traditional search is only retrieval, a porter of information,” she said. “But a probabilistic model trained on massive data outputs results after internalizing information; it is an alchemist of information.”

Producer Xiong Xiaoling and her team were early users of Script Workshop while preparing the idol drama “Love You,” which aired in February 2025 and reached the top spot on iQIYI International’s popularity rankings in 12 countries and regions. She said the episode plot-point function was particularly useful for proposals and communication with partners. To develop the evaluation function, the algorithm team studied how Zhang’s research team assessed scripts and translated their criteria into machine-readable dimensions. A comment such as “too many threads,” for example, corresponds to dozens of pieces of information the model must extract and summarize: how many plotlines a script has, whether they are connected or isolated, when they converge and whether flashbacks or nonlinear storytelling have fragmented a clear main line. The team turned prompts into AI agents. By the end of 2024, Script Workshop had provided evaluation and speed improvements for 5,000 novels and scripts. In April 2025, it integrated Chinese models including DeepSeek to improve understanding of Chinese language and culture, and it can evaluate a 500,000-character script in under 30 minutes across seven dimensions. Almost all studios now use it.

Image Workshop, an AI-based platform for full-process image and video generation and management, generates storyboards from scripts, as well as concept posters, character designs and scene designs. With models such as Midjourney V6.1 and Stable Diffusion XL, text can be turned into visual assets, speeding up early art concept design. At iQIYI World Conference, Liu cited the character Duan Banxia in the drama “Wuyou Du”: Script Workshop extracted the textual description of the character, and Image Workshop generated an ancient Chinese-style image.

Editor's Summary

iQIYI has spent 580 days building generative AI into both viewing and production. Its Skip Watch feature compresses long-form drama, while Script Workshop and Image Workshop are used in script evaluation, story development and visual design. The company says most employees have used large models and that its tools have evaluated thousands of scripts, while it has avoided simple consumer AI applications.