Writing a prompt that survives reruns
Three habits that take you from lucky-once results to a prompt you can actually iterate on.
AI image and video models are probabilistic. Run the same prompt twice, you get two different outputs. That’s fine for discovery; it’s painful when you have a result you like and need a small tweak.
Three habits that move you from lucky-once to controllable:
Pin the seed. Every model on VideoGen exposes a seed param (default -1 for random). The moment you have a result worth iterating on, copy the seed and lock it. Now you can change one phrase at a time and actually see the effect.
Front-load the subject. Models give more weight to the start of the prompt. Lead with what the image is of (subject, framing, lens), then layer in style and lighting, then negatives. Buried instructions tend to lose to the words before them.
Use the negative prompt for video. The video models on the catalogue accept a negative_prompt param. It’s the cleanest way to say “keep the composition stable” — list the kinds of motion you don’t want (sudden zoom, body translation, background swap) instead of trying to encode their opposite in the positive prompt.
Try it on a fresh studio session — pick a model, pin the seed, and iterate from there.