Dalian Works
aifilm

Why the Finished Videos Have No Real URL in the Database and Live Entirely in the Jimeng Cloud

Published 2026-11-03 · Translated from the Chinese original

When you browse through the database records of the final film in the Yuncao Academy system, the most convention-defying fact is that the vdurl field in the dbvideo_process table does not store a downloadable address of the video file at all, but rather a business identifier string from the Jimeng platform, in a format resembling dream.xxxx-xxxx-xxxx, or a plain text string containing a task ID. Across all 1,029 storyboard records in the entire database, not a single vdurl points to a real video file address that begins with http and ends with .mp4. If you try to open the contents of these vdurl fields directly in a browser, all you will see is the task detail page of the Jimeng cloud, and the page itself does not provide video streaming to non-logged-in users. This fact leaves uninformed developers bewildered when reviewing the data, and some even mistakenly believe that all film assets have been lost. In reality, the video file entities are stored in the cloud storage bucket of the Jimeng video generation platform. The Yuncao Academy system only interacts with Jimeng through APIs and records a traceable task reference identifier locally. The root cause of this design lies in the invocation model of the Jimeng video generator. When the client calls the Jimeng interface to submit a video generation task, the parameters passed in include the first frame image path, the last frame image path, and a structured aiwords prompt string. After successfully receiving the task, Jimeng returns a task ID and binds that ID to the task metadata under the user's Jimeng console account. After the video is rendered in the cloud, Jimeng does not actively push the video file to the client. Instead, it provides an interface for querying task status. The developer polls the status using the task ID, and when the status becomes completed, Jimeng exposes the preview address and download address of that video. However, this download address typically carries a temporary signature and validity period, generally lasting only a few hours. Moreover, the Jimeng platform's contract logic defaults to retaining only the last thirty days of generation records, and video files beyond that period are automatically cleaned up. Therefore, during the film production process, the Yuncao Academy does not proactively store temporary download addresses in the database, because once they expire, they would dirty the data and in turn interfere with the production process's judgment of whether video generation succeeded. In the actual production of "The First Avenue," all video generation tasks were pushed to the Jimeng cloud for completion. Since the number of tasks reached 1,029, and each video took anywhere from forty-five seconds to three minutes to generate, if every temporary download address were saved locally, the storage volume would be small, but the validity of the addresses would be extremely fragile, and they would all expire within a few days. To manage this uniformly, the system adopted a more pragmatic strategy: store the Jimeng task ID in the vdurl field, and store the completion status of the task and the error code returned by Jimeng in the status field. When a record's vdurl value begins with dream., it indicates that the task has been successfully submitted. As long as the status is completed, the video segment can be viewed and downloaded. However, if the entire video segment needs to be viewed again or extracted, the operator must log back into the Jimeng console or initiate an export request in the Yuncao Academy, from which the backend calls Jimeng's download interface to obtain a new temporary link. This also explains why there is no real video URL in the final film data: the system itself does not assume the responsibility of storing and distributing video files. Some might ask: since the videos are all in the Jimeng cloud, if the platform fails or the account is banned, wouldn't the entire film vanish into thin air? This concern is reasonable, but the Yuncao Academy's coping strategy is to guarantee the completeness of metadata for every video generation task. The Jimeng platform provides an interface for downloading the original video for each task. After the Yuncao Academy detects that a task status is completed, it proactively calls that interface within seven days to download the video file back to the local server, saving it to the videos folder under the bookcode directory, with the file named by scenecode. However, in the production workflow of "The First Avenue," the final decision was not to download to local storage, because the hard drive space on the three local computers was limited, and the 2,058 first and last frame images had already occupied about 30 GB of storage. Saving an additional 1,029 MP4 files with an average size of 60 MB each would bring the total capacity to over 60 GB, putting pressure on the synchronization and backup of the three computers. Therefore, in the end, only the Jimeng task IDs and generation information were retained. The complete videos were stored directly in the Jimeng cloud and viewed through streaming playback on the Jimeng platform's preview page. When the author himself needed to deliver the finished film, he used the merge-and-export function in the Jimeng console to concatenate all video segments into a complete film in one operation, then downloaded it locally for final editing. What needs to be made even clearer is that the vdurl field has not become insignificant just because it stores a task ID. It remains the key basis for troubleshooting and data tracing throughout the entire system. Suppose a shot's video is found to have a visual anomaly during final merging, such as a black screen on the last frame. The operator can use the task ID corresponding to that shot's vdurl value to query Jimeng for the specific parameters used in that video's generation, including the first frame image, last frame image, prompt template, and the error log returned by Jimeng. Based on this information, one can precisely determine whether the issue was a placeholder problem in the secondpic field of the image generation stage or a defect in Jimeng's own algorithm during generation. By contrast, if the system forcibly stored download addresses in the database, once they expired, the address could no longer locate the generation record, and the tracing value would be lost. Therefore, storing the task ID in vdurl is meant to preserve generation records over the long term, while the temporary download address is considered data suitable for immediate use and not requiring long-term storage. In later versions, the Yuncao Academy provided a scheduled batch download feature that converts all task IDs contained in the vdurl field into long-term storage links for Jimeng-generated videos and periodically syncs backups to Alibaba Cloud OSS. But this is merely icing on the cake and does not affect the existing workflow. Additionally, the Jimeng platform has a clear file lifecycle management rule for each user. Generated videos are retained for thirty days by default, and users can proactively extend the retention period in the console or download them locally. The reason the Yuncao Academy was able to reference these video data stably within a month is that during production it called Jimeng's download interface in batches to obtain the original videos and cached them locally. The final delivered film was derived from these cached files, not from direct online previews. After the film was completed and edited, the cache could be selectively deleted to free up space, while the vdurl in the database was retained purely as metadata. This strategy ensures that all video files can be re-extracted when needed, while daily operations do not suffer errors from expired addresses. For anyone wishing to replicate this methodology, the core understanding is not to treat the vdurl field as an ordinary file path. It is merely a cloud task identifier, and a truly efficient workflow relies on on-demand query and transfer mechanisms. When you see in the database that VDURL is empty or an unremarkable alphanumeric string, do not mistakenly conclude that the task failed. Instead, you should query the status field to confirm whether the video was successfully generated, and use the task ID to log into the Jimeng platform to view the actual video content. This is the most counterintuitive yet most practical point of the entire system: data storage is not the same as file storage, and final film management under cloud-based video production requires redefining the meaning of data fields.