How do you iteratively improve AI video output by telling the tool what you don't like?

You improve AI video output by running generation as a conversation, not a series of one-shot prompts: pin your best generation, state specifically what's wrong and what should replace it, and keep iterating in the same session — the tool internalizes your corrections, so later shots in the scene land faster and more accurately.

Iterate in a conversation, not in isolated prompts. Keep every correction in the same session so each new generation builds on the last — one filmmaker who documented this workflow describes it as "literally like just having a conversation back and forth until you get what you want." The invideo agent works this way natively: it holds the full back-and-forth in context, so you never restate the whole prompt to fix one flaw.

Pin the generation that's closest and request targeted variations. When an output is 70% right, don't regenerate from scratch — pin that specific video and ask for another one like it, but with X, Y, and Z changed. Naming the exact dislikes ("same shot, but remove the camera drift and darken the water") keeps everything you liked while surgically replacing what you didn't.

Phrase dislikes as replacements, not bare negatives. Models handle negative phrasing inconsistently — some treat "no fog" as a cue that fog is relevant. Convert every dislike into a positive description of what you want instead: "the lighting is too flat" becomes "harder side-light with deeper shadows." This holds across Veo, Kling, and Seedance 2.0, all of which run inside invideo, so you apply one feedback discipline while the invideo agent routes each shot to the right model.

Anchor feedback to a visual reference when words fall short. If you can't articulate the dislike, upload a still from your own footage as the reference — color, contrast, and look get corrected against something concrete rather than another round of adjectives.

Let your corrections compound into a visual language. Every like/dislike you register teaches the session your look, tone, and effect — in one documented production, after the first shot was refined through feedback, the second shot in the scene came out correct on the very first generation. As the filmmaker put it: "when you establish what the look, the tone, and the actual effect is, when you move on to another shot in the scene, it does it a lot quicker and it's a lot more accurate because it's learning exactly what you want and how you like it."

Know when to stop iterating and start assembling. First-generation output is rarely final, and some flaws aren't worth another round of feedback — extract the best moments from several imperfect generations and splice them into one complete sequence instead of chasing a single perfect clip.

Watch some of these to see what works for you:

AI is so new. It has been around just for a couple of years now and it is constantly learning, it's constantly changing and evolving, and you have to let it know what you like and what you don't like.

— Alex Arfaoui, independent filmmaker