QCon Live Discussion Explores Stable AIGC Delivery in Production; 2026 Shanghai Event to Feature AIGC Content Industrialization Track

The discussion began with a seemingly simple question: what has been the biggest change in AIGC content production over the past year. The answers from the three frontline leads pointed in the same direction—the unit of measurement has changed. Lu Xiangju said the biggest change is that AI has moved from the Demo stage to the stage of stable delivery and practical works, and AI is moving from short videos and short clips into the era of professional works. Zou Juan summarized this change as the unit of competition shifting from a single clip to a complete production pipeline; the finished product may be a program as a video file, or a real-time pipeline work such as live streaming.

Zheng Zhong added background from the supply side: over the past year vendors have released three to four hundred or more models, and competition has shifted from "who can make a better Demo" to three key questions—whether creativity can be converted into a production process, whether the consistency, stability, and controllability of complex content can reach industrial grade, and whether it can go into production and how much automation is possible. He also mentioned that customers no longer care only about model capability, but about whether a set of high-quality finished videos can be delivered, and whether the production pipeline around the finished video can be implemented and scaled.

This set of judgments means the coordinate system for evaluating AIGC has switched from "single-point quality" to "production tasks." Zou Juan mentioned that people used to be easily impressed by a 10-second video, but now that we have entered real industrialized scenarios such as short dramas, marketing, and e-commerce, users will ask: can the characters stay consistent? Is the product deformed? Can it be delivered on time? Can the cost be held down?

One correction that was repeatedly emphasized in the discussion concerned how cost is calculated. The discussion proposed that the real cost is not the price of a single generation, but the sum of generation, retries, review, manual editing, quality inspection, storage, rendering, operational coordination, and other costs divided by the number of usable qualified finished videos—far higher than "how much for 5 seconds of one model call."

The reason this formula matters is that it directly translates the randomness in model calls into money. Zheng Zhong said a stable model might be 2 draws for 1 or 3 draws for 2, while an unstable model might be 10 draws for 1 or 20 draws for 1, and since model or workflow costs are charged by number of calls, this brings a large amount of cost. He further pointed out that uncontrollable output also brings two other production problems: it cannot be automated or turned into a process, and SOPs cannot be accumulated; and content quality varies across stages, so it cannot be integrated into a complete high-quality piece of content.

This leads to the relationship between "stronger" and "more controllable." Zheng Zhong believes controllability itself is an important dimension for evaluating whether a model is stronger; a model that stably meets requirements eight or nine times out of ten is more suitable for production than a model that is stunning once out of ten. His judgment is that if a model produces only one stunning result out of ten draws and nine are unusable, it is essentially still hand-crafting single models and exploring capabilities, and is not suitable as a productivity tool.