We have 12 videos to deliver this weekend. Lol. There's always joy in bringing ideas to life. It feels like you're in motion, and it gives you purpose.

The current AI race This is so true, I’m hoping Gemini will make a great come back

I get asked this almost every time now. How do your AI videos sound African? The answer is annoying, free, and takes about two minutes to set up. But first, let me name the actual problem, because I don't think enough people are saying it out loud. The visual side has come a long way. Characters look African now. Skin tones read correctly, features are right, wardrobe is believable. Two years ago that was a fight. The voices have not caught up. Not close. Go to ElevenLabs right now and search for Nigerian voices. What you'll find, mostly, are UK-based voice actors with Nigerian nationality. Technically Nigerian. They do not sound like anyone I know. There's no accent in there, or there's a version of the accent that has been sanded down for a British ear. That gap tells you exactly what data these models were trained on, and it wasn't ours. Here's why it matters more than people think. A Nigerian audience will forgive a slightly wrong shot. They will not forgive a wrong voice. The moment a character opens their mouth and sounds like a BBC presenter doing an impression, the whole thing collapses. You lose them in one second, and you don't get them back. So this is how I've been getting around it. I've been testing this in Seedance 2.0 and 2.5. The first part is the one that actually does the work. I use real voices. If I need a specific kind of voice for a project, I go to someone I know who has that voice, and I ask them to record it. Five to fifteen seconds is enough. I explain clearly what it's for, that it's an AI project, and I get their consent before anything else happens. That part is not optional, and it's not just an ethics thing. If the work is commercial, you need to be able to say where that voice came from. Then I save that recording as a voice reference and feed it into the generation. The second part is smaller, but it matters. When I write the prompt, I specify the accent explicitly. Not "he speaks." I write "Nigerian-accented English." If you leave accent out, the model defaults to American. Every time. It has no reason to guess otherwise, because that's the middle of its training data. You have to state it. I want to be clear that the prompt line alone won't save you. I've tested that. Writing "Nigerian accent" with no reference gives you something vague and often embarrassing. The reference audio is doing the heavy lifting. The prompt is just making sure the model doesn't drift away from it. A few things I've picked up doing this. Record clean. A quiet room and a phone is fine. Background noise gets baked in, and you'll hear it in every generation afterwards. Have them speak naturally rather than performing. A normal conversational read carries the accent more honestly than someone putting on their announcer voice. Match the energy to the scene. If the character is angry in your film, a calm reference will fight you the whole way. And keep a library. I've started saving references the same way I save character sheets and wardrobe. Different voices, different registers, all labelled and filed with the consent on record. Once you have five or six good ones, you stop starting from zero on every project. Here's the broader thing I keep coming back to. Nobody is coming to fix this for us. These models get trained on whatever data is easiest to gather, and our voices are not in that pile. Waiting for ElevenLabs to add a proper Lagos accent is not a strategy. The workaround is the work right now. And the creatives who build their own reference libraries are going to have something no model release can hand to anyone else. If you've found a different method that works, I genuinely want to hear it. This is the part of the pipeline I'm still least happy with.

Nine tips I wish someone had told me before my first AI film. Most of them cost me money to learn. These aren't theory. Every one of these came from a project where I did the opposite and regretted it. 1. Sort your assets before you generate anything. Voices, characters, locations, props. All of it, locked and organized before production starts. Not during. Not "I'll figure it out when I get there." The moment you start generating without your assets sorted, you're paying credits to make decisions you should have made for free. 2. Specify the accent in your prompt. This one is small, and almost nobody does it. If you don't tell the model which accent your character has, it will default to a generic American accent, and your film will quietly lose its sense of place. Write the accent into the prompt. Every time. 3. Say no to background music in your prompt. Models love adding their own music. It's almost always wrong, it's almost always generic, and it fights whatever you actually plan to score the scene with. Specify no background music and take that control back for the edit. 4. Use grey backgrounds for your character sheets. Neutral grey. It isolates the character cleanly, keeps the model from picking up environment colour into the reference, and makes those sheets reusable across completely different scenes. A small technical choice that saves you constantly. 5. Avoid flashlights. Use overhead light sources. Flashlight lighting in AI video looks wrong almost every time. Harsh, unstable, physically confused. Overhead lighting reads as natural to these models and gives you far more consistent results. Choose the lighting the model can actually handle. 6. Keep fewer people in the frame. Every additional person in a shot is another chance for the model to break. Faces drift, bodies merge, hands do things hands don't do. Fewer people means cleaner generations and fewer wasted credits. Write your scenes knowing this. 7. Write your own script. This is the one I feel strongest about. Write it yourself. If screenplay format isn't your strength, that's fine; write the full story in your own words and let AI convert it into screenplay format. But then read it. Properly. Understand it before you generate a single frame. The story has to be yours, or the film ends up belonging to nobody. 8. Seedance 2.5 is quietly a great editor, not just a generator. If you can afford it, it will rescue shots that came out badly, and it's genuinely strong at maintaining continuity from a video reference. Most people are using it purely to generate. It's also a repair tool. 9. The real secret is the idea, and watching everything. The idea has to come from you. That's what gives you control over your shots, your pacing, your decisions. Without your own idea driving it, you're just accepting whatever the model hands you. And the fastest way to develop that? Watch a lot of films. Watch a lot of art. Study what other people have already made and why it works. Everything I know about framing and rhythm came from watching, long before it came from any tool. The tools keep changing. The craft underneath them doesn't. Which of these do you already do, and which one are you guilty of skipping? I skipped number one for months, and it cost me.