
Verify the reference binding in your saved request, then inspect the completed video's opening, middle and ending against an approved visual brief. A successful task proves neither that the intended image was selected nor that every subject detail survived. Keep attachment correctness separate from the model's creative interpretation.
Key Takeaways
- Trace the submitted reference to its actual approved source, not a thumbnail label.
- Evaluate the chosen image role using the intention you approved.
- Review complete motion and sound before accepting the clip.
Preserve a source-to-video review record
Your application should retain the source selector, approved source identity, image role, submitted prompt and returned video task ID. This example is review metadata, not a generation payload:
{
"shot": "window-turn",
"sourceSelector": "image_task_id",
"approvedSourceTaskId": "SAVED_IMAGE_TASK_ID",
"imageRole": "first_frame",
"videoTaskId": "SAVED_VIDEO_TASK_ID",
"mustPreserve": ["navy jacket", "single visible person"],
"intendedAction": "turn gently toward the window"
}
The IDs are placeholders. The Imagild API guide requires a usable same-account source: an uploaded file_id or a successful image image_task_id, never both.
Inspect the attachment before blaming generation
Open the source associated with the saved operation. Confirm that it is the approved image, not another image that finished more recently. Reordering UI cards or replacing a draft source can otherwise hide the real binding.
The reference attachment audit explains that class of mistake. If the wrong image was submitted, correct your application's reference mapping before paying for repeated creative prompt changes.
Keep the accepted video's record immutable while repairing the draft. A new source and new generation are a different operation, not an identical retry of the previous one.
Review the role against its intended use
With first_frame, compare the opening composition with the starting scene you intended to animate. With reference_image, compare the important reference traits with the requested newly composed shot. The roles express different input intentions; neither guarantees exact pixels or identity throughout the clip.
Create concrete acceptance checks before viewing the result. “One person remains visible in the navy jacket” is reviewable. “Looks like the source” can hide different expectations about face, outfit, camera angle and background.
The reference scope guide helps decide which parts of the original scene matter. Do not label a deliberate new background as a binding failure if that was the approved brief.
Inspect motion, media and soundtrack
Check the opening, peak action and final frames, then play the whole file at normal speed. Sparse frame samples can reveal a defect but cannot prove that every intervening moment is clean.
Review unintended extra people, disappearing objects, distorted movement and abrupt endings. Listen for unwanted dialogue, music or sharp audio changes rather than assuming the clip is silent.
External media tools such as ffprobe can inspect streams and duration. They cannot establish whether the person or action matches your brief, and they are not an Imagild visual-QA endpoint.
Choose the correction that fits the evidence
If the saved source is correct but the requested action is ambiguous, revise the motion brief. If the content is usable except for a short ending, evaluate an external edit. If a reference cannot be resolved, fix that dependency rather than changing model settings at random.
A higher tier or enabled enhancement does not guarantee preservation. Use the model hub to quote any deliberate new generation separately.
Pass the shot only when its binding is traceable, the actual downloaded file satisfies the visual brief and the soundtrack is acceptable. Record a failed review as a failed acceptance decision, not as evidence that the API request itself was invalid.
Related tasks
- After Reordering References, Does “Image 1” Still Point to the Right File?
- Should an AI Reference Replace an Object or Bring in the Whole Scene?
Sources and scope
- Imagild API console: Actual request and task envelopes, strict parameters, references, membership, points usage and authenticated result retrieval.
- Imagild model hub: Current advertised model capabilities, resolutions and generation point rates.
- FFmpeg ffprobe documentation: External stream and duration inspection, distinct from visual semantic review or a native Imagild QA endpoint.
Product instructions were checked against the current Imagild API contract and the public model catalog. Code, budgets and acceptance checks are illustrative; this article does not claim a new paid generation test, an uptime measurement, a search-volume study or a ranking of AI citations.
Prepared by Imagild Editorial, the company editorial team responsible for these guides. AI assisted research and drafting. This is documentation-based guidance with proposed examples, not a claim of hands-on testing or an independent product review. Report a correction.
Cover: Original Imagild product screenshot captured on 28 October 2026; model availability and prices may change. This image is not an API response or a performance benchmark.