If your brand runs on Meta ads, you already know the first script is almost never the winner. It's an iterative process where you find the winner by testing, which means many scripts, many hooks and many versions of the same narrative running at the same time.
Doing that with a real shoot means casting, locations, a crew and a lot of managing for every new version. Doing it with AI is faster, but only a little. You still write prompts for the characters, then the environment, then the video, often across three or four different platforms, and then you do it all again for the next script.
That repetition is painful and often becomes the bottleneck. It is also the part our team at Chromatic AI has solved.
Three artists worth watching this week
Video models this week
Seedance 2.5 - the long single-take. Up to 30 seconds in one native pass at full 1920x1080, no upscale step, and up to 50 references per generation (30 images, 10 video, 10 audio) against 2.0's nine images and three clips. New in 2.5: a single voice, music or sound-effect track drives pacing, beat-matching and lip-sync across 10+ languages. Region-level editing swaps a subject or replaces the music while the rest of the frame stays untouched. Reach for it when the brief is one continuous 20–30s spot.
MiniMax H3 - the reference-heavy multi-shot. 5 to 15 seconds at up to 2K with native stereo audio, and one generation takes up to 9 images, 3 video clips and 3 audio files. It reads each for a different thing: who the character is, how they move, how the camera behaves, what voice comes out. It also does multiple shots inside one clip, which is why our jewellery sequence came out of it at a fraction of Seedance's cost and turnaround.
Kling 4.0 - the keyframe-steered long take. 3 to 30 seconds in one pass at up to 4K, double Kling 3.0's 15-second ceiling, with 10-bit HDR on the way. You can pin up to 10 keyframes along with first and last frame, and it builds one continuous shot through every beat you set. One generation holds up to 15 references across images, video clips, saved elements and up to three voices, with stereo audio and lip-sync across languages, accents and dialects. Finished clips extend to 2 minutes and keep the same character and setting. Reach for it when the story has fixed beats it has to land in order.
Image models this week
GPT Image 2.5 Sunburst - the precision tier. Built for premium work that needs tight control across edits, with strong subject preservation from reference photos, output to 3840px and transparent backgrounds. Changes hold across a long edit chain instead of drifting. This is the campaign-creative and product-shot model.
Nano Banana 2 - worth a naming note in the issue, since most people still say Nano Banana Pro and the board lists this as gemini-3.1-flash-image. it is what our image skills are built around. Speed and reference handling are why it survived our pipeline while higher-ranked models did not.
GPT Image 2 (medium) - the settled one. When you need a result you can predict rather than a result that might be better, this is the one.