What a Tiny Vampire Bat Taught Me About AI Storytelling
Lord Waffles the Third was only supposed to appear in one 15-second video. Instead, he became the tiny star of an increasingly elaborate gothic comedy series, acquired a devoted vampire father named Count Fierce, and somehow turned me into a full-time bat continuity supervisor. Along the way, he taught me something genuinely useful: AI storytelling is not really about making a perfect video. It is about creating a character people want to see again.
One Image Was Enough
I create my videos in Runway using Seedance 2.0, and Lord Waffles has exactly one character reference image. That’s it. No elaborate character sheet. No 12-angle turnaround. No folder labeled “Lord Waffles Expressions Final FINAL 2.” Just one clear image that shows what he looks like. The prompt handles the rest by reminding the model that he is a realistic tiny bat with dark fur, natural wings, large ears, and no bizarre fantasy anatomy. This does not guarantee perfect consistency, because AI occasionally looks at “tiny bat” and thinks, “Excellent. Here’s your goat.” Still, one strong reference image combined with consistent instructions has been enough to make him recognizable from video to video.
Consistency Comes From Rules
Lord Waffles and Count Fierce live in a modern, high-budget, live-action gothic comedy. It should not look animated, magical, hazy, vintage, or like anyone is about to burst into a musical number. I repeat those rules in almost every prompt because the model does not remember the series the way the audience does. Every generation must be reminded who these characters are and what their world looks like. The emotional rules matter too. Count Fierce is elegant, dramatic, and completely sincere. Lord Waffles is tiny and quietly tolerates whatever Count Fierce has planned for him. That relationship does most of the storytelling for me.
Fifteen Seconds Can Take All Day
A single 15-second Seedance 2.0 generation at 720p in a 16:9 format takes about an hour. At the end of that hour, there is absolutely no guarantee that the video will be usable. A prop may disappear. An extra hand may enter the scene. Someone may suddenly turn and stare directly into the camera like they have just become a reality TV star filming a confessional. When that happens, I revise the prompt and try again. Some videos take 15 to 20 prompt revisions before I get a result that works. Yes, that means 15-20 hours of video generations down the drain. It can be very frustrating.
The First Prompt Is Only a Draft
I no longer treat the first prompt as final instructions. It is more like opening negotiations. Every failed generation reveals how the model misunderstood the scene. I clarify the action and remove unnecessary details. Sometimes the model needs more detail. Sometimes it has too much detail. AI video creation currently feels like being the writer, director, continuity supervisor, costume designer, vampire wrangler, and exhausted parent of a very small fictional bat all at once. A large part of the skill is not simply knowing what you want. It’s learning how the model is likely to misunderstand you.
We Still Have a Long Way to Go
AI video has improved dramatically, but it is still far from finished. One of the biggest problems is what happens when a generation almost works. Maybe 13 seconds are perfect, but the bat doubles in size at the end. Maybe the performance is great, but Count Fierce suddenly grows a third hand.
Ideally, I would be able to select that exact moment and fix only the problem. Current editing tools offer some control, but not enough to reliably repair small continuity errors without changing the rest of the video. I do not just want longer clips or higher resolution. I want precise editing. I want to keep the parts that worked, isolate the one cursed second, and repair it without starting over.
We are not there yet.
