Ruairi Robinson directs an epic, AI-generated promo for David Guetta's team-up with Dimitri Vegas and Loreen, showing the fight against evil in a post-apocalyptic world.A cinematic trip from start to end, the video for Pum Pum is one of the most realistic AI videos we've seen so far, helped by the quick-cut, frantic aesthetic and dystopian visuals. Though close-ups and lingering shots on faces still give an uncanny valley vibe, it works with this concept, and is a sign of how far the technology has come in just a few short years. "When I first heard the song, the first spark of the idea was an epic post-apocalyptic motorbike race, then I threaded in a Kill Bill-style revenge story and cranked everything up to ridiculous levels: revenge, robots and epic action," says Ruairi Robinsin "Basically, a movie for my boyhood self. Ie Akira, The Road Warrior, Ghost in the Shell, and Jurassic Park all thrown into a blender."Not ripping them off, but chasing that same pulse I felt watching them as a kid. The next challenge was turning Dimitri Vegas, Loreen and David Guetta into believable photorealistic characters in this world: something most AI models are terrible at. I needed them to look consistent, realistic, expressive and able to sing. I tried just about every AI tool in existence, and quickly learned a useful rule: the more loudly a model is hyped on Twitter, the less helpful it tends to be in production."Most commercial closed models fall apart when continuity’s involved. If you have a motorbike turning around, the model doesn’t know what the back looks like and it just invents it differently every time. So for a lot of shots I turned to a Chinese open-source model called WAN, which lets you train almost anything the base model can’t handle. The artists sent me selfies with every expression imaginable, and we collaborated with the Conductor software platform that connects multiple AI tools through a proper render queue. This helped coordinate a small team to handle training. Utilising various cloud systems we tested different setups to see what worked best. Vehicles started as single renders, then I used additional tools to generate alternate angles and lighting, building datasets to train LoRAs, so I could drop those vehicles into any shot."The catch with LoRAs is “concept bleed.” If I trained Loreen in battle armour next to a robot, they’d start merging. For example, Loreen with cables and the robot with long hair, a sort of 1980s Cronenberg mutation. To fix that, I used a WAN tool, which swaps trained characters into existing scenes. It sounds straightforward, but nothing in ComfyUI ever is; it’s Houdini for AI — a tangle of spaghetti nodes that somehow works, sometimes."