Guide
Video Generation and Retry
How to use video models, reference images, first/last frames, and failure retries
Video Generation and Retry
Ordinary Reference Images
Ordinary reference images describe characters, scenes, clothing, or visual style. You can reference them from the asset library or upload them temporarily in the current input box. Before submitting, confirm that the image is clear, its content matches the request, and the selected model supports reference images.
First and Last Frames
First/last-frame mode requires specifying the first frame and the last frame separately: the two images cannot share the same character slot, nor can they be submitted again in the ordinary reference area. If either frame is missing, the platform blocks the submission before creating the task.
Multiple Images and Reference Video
When you need "multiple reference images + one reference video", add each image and the one reference video separately as reference assets. There is no separate "deep video" field for reference video in the API; submit it as an ordinary reference video. For capability boundaries, ordering rules, and audio notes, see Multiple Images and Reference Videos.
Model Selection
After you explicitly select a video model, the workspace uses that video capability to create the task. When smart planning is used, the system first organizes the requirements for this creation; when it is not used, confirm yourself whether the model, ratio, resolution, and duration suit your goal. Whether a model supports reference images, reference video, and audio is subject to the model details.
Reference Image Read Failure
When you see "failed to read reference asset", it means the model service cannot read the reference asset; this is usually not a prompt problem. Handle it in the following order:
- Confirm on the asset card that the status is "Available"; do not just check whether the local thumbnail displays;
- For portrait assets, click "Refresh" and wait for the material processing to complete;
- For asset library references, use a registered asset; for temporarily uploaded assets, select the original file again before submitting;
- If the same asset keeps failing, switch to a compliant, readable image, or contact platform support to check the material permissions.
If the task has been submitted to the model service, the workspace shows the status or error actually returned by the model. The asset status check before the creation request does not produce a video task; only after the task is successfully accepted does it enter the waiting and result query stage.
Direct Retry
"Direct Retry" keeps the original conversation and original parameters, resubmits in place, and refreshes the waiting status. It does not restore the input box, create duplicate messages, or treat the old failed result as a new request. A retry produces a new task and new charges.
Saving Results
Successful videos are saved to the current creation record; you can download them, copy the public prompt, or reference them in later creation. Failed results do not generate downloadable placeholder files; it is recommended to transfer download URLs of generated results to your own storage as soon as possible.
Pre-Submission Checklist
| Item | What to check |
|---|---|
| Model | A video model that supports the current creation method has been selected |
| Prompt | Includes the subject, action, scene, and necessary shot requirements |
| Reference assets | They show as available, or the temporary files for this submission can be read normally |
| First/last frames | Both character slots are filled, and they are not the same image |
| Ratio and duration | Within the ranges the model supports |
| Retry after failure | If the cause has not changed, fix the assets or parameters before retrying |
