Turn your vague “coffee video” idea into a structured video prompt. Return JSON: {"prompt":"one subject and one clear action","anchor":"description of the anchor starting frame that fixes the composition and color","motion":"what kind of movement is in the frame (light, one)","duration":"duration in seconds, briefly"}
The generator does not draw the entire “video” - it either unfolds frames from noise based on text, or animates the starting picture, predicting the movement of pixels between adjacent frames. Hence the main difference from statics: time and movement are an extra dimension where “drift” easily accumulates (the face floats, the object is deformed). Control levers: short clip (3-6 sec), one subject, one clear action. Anchor frame - the starting image (from the picture mode) - fixes composition, color and face in advance, reducing unpredictability. Insider: first take a strong static shot, and then “animate” it - this way you control ~80% of the result even before movement. Second technique: make the movement light and single (steam rises OR the hand pours - not all at once). A typical rookie mistake: cram three actions and an “epic camera” into one video - the model will smear everything into mush and produce a flickering defect.