Troubleshoot poor AI image and video results

Sort the issue into task state, input, subject, composition, text, visual treatment, action, camera, or output file. A failed task needs a different response from a successful render that misses the taste target. Misspelled package text will not be fixed by changing a video motion value. Write one observable statement such as “the product label changed” or “the face distorts at three seconds.”

Avoid keeping only notes such as “bad” or “not premium.” Convert them into differences you can inspect: the subject is too small, the background hue is wrong, the camera moves left instead of forward, or an arm breaks during the turn. Location and timing make the next test specific.

Confirm whether you selected text generation, editing, image-to-video, or another currently available workflow. Check that the upload control shows the intended start frame, end frame, subject, or style asset. Inspect orientation, sharpness, subject edges, and crop. If references disagree about identity, clothing, or color, remove the weaker one.

Read the current duration, ratio, resolution, and other controls again. A setting exposed by another model may not apply here. When the source is blurred, tiny, or covered by interface elements, a longer prompt cannot restore information that is not present.

Reduce the job to subject, one main action or change, setting, output form, and one critical constraint. Remove long style lists, conflicting camera directions, simultaneous transformations, and broad negative lists. Verify “keep the product and label; replace only the background with light gray” before adding wet stone, rim light, and mist.

If the minimal version fails in the same place, prompt length may not be the cause. Keep that baseline while comparing media, settings, or models. If it succeeds, add details one by one until the instruction that introduces the drift becomes visible.

For identity drift, start with one clearer reference where the subject is larger and its outline is complete, and reduce simultaneous changes to styling and scene. For composition, change only viewpoint, shot size, subject position, and target ratio. Do not replace face, clothes, background, and camera position in the same test.

Exact text must be readable in the source. State its content, position, and preservation requirement, but review every result manually. Hands, jewelry, package corners, trademarks, and thin structures benefit from a simpler frame and larger subject. Important commercial work should not be published without inspection.

Determine whether the video problem belongs to the subject or the camera. Walking, turning, and cloth movement are subject actions; push-in, pull-back, pan, and orbit are camera moves. Keep one principal action and one simple camera instruction, with direction, speed, and range. A short clip cannot reliably carry many scenes and transformations.

If a face or limb becomes unstable at one moment, reduce the motion range, narrow the goal, or use a start image closer to the intended pose. If the camera jumps, remove contradictory directions. Treat the current form as the authority on dedicated motion controls.

A useful comparison holds the prompt, media, and target constant. Change one clear setting within the same model first. If the same defect remains through two or three rounds, compare another model that supports the same type of input. Do not switch model, reference, ratio, and prompt together.

Record identity retention, composition, text, action, camera, credits, and whether the result is usable rather than selecting only the most striking first impression. For quality-sensitive work, run a small sample before expanding a batch.

Continue controlled refinement when the task succeeded, the problem is content-related, and each test changes an explainable variable. Retry with corrected input after a confirmed failed task and visible credit return. Stop creating jobs when the task is active, upload is abnormal, the same input repeatedly fails, or status and credit records disagree.

For support, keep the task identifier, time, model, mode, visible error, setting summary, and a screenshot without sensitive material. Do not delete the failed record or publish private media. A concise account of expected result, actual symptom, and variables already tested is much easier to investigate.