After Vidu’s multi-reference Q1 went live yesterday, Kling’s multi-reference feature will likely be updated soon as well. I’ve always believed that multi-reference is a crucial method for solving character consistency and scene consistency. This update is the first step toward completely resolving consistency issues, and within a few months, the consistency problems that plague most creators could be resolved.
Although multi-reference might still have some character proportion issues currently, soon, we’ll only need to create frontal character designs and scene designs, then let multiple characters perform in these scenes. The entire production approach for AI videos will also change, evolving from image-to-video control methods to direct video generation methods.
What will the evolution of multi-reference bring us? It’s further democratization of creation – the barrier to producing long-form narratives will be lowered once again. Because consistency is the foundation of long-form storytelling. Style consistency, character consistency, and scene consistency – at least the latter two are no longer worth obsessing over.
I’ve been working on character consistency since the SDXL era and still have character images from that time. Then came Flux’s character LoRAs – every time I made a film, I had to first train a set of character LoRAs. Will the multi-reference update render LoRAs unnecessary? In most cases, yes. What remains are some shots that multi-reference can’t handle.
Will I still train LoRAs? Actually, I’m already doing one-shot LoRA training now. So it’s consistent with multi-reference needing just one image. When creating character design, you typically need at least one full-body shot from the front anyway. The technical skills and aesthetic sense from character design work are still preserved.
So here’s the question: what should we focus on after we no longer need to obsess over character consistency?
I know what I can still focus on – after not obsessing over character consistency, I still need to focus on IP consistency. That is the interconnectedness between different stories. Every story exists within the same worldview, with a consistent IP tone, ultimately forming multiple stories into an IP matrix.
This is the first thing that came to mind – a type of consistency that not everyone can achieve. This requires a certain level of story and character design. It also requires designing a logically rigorous fictional world. This suits me well, but it may not be suitable for everyone. For instance, many sci-fi stories can simply take place on Earth and Mars.
Looking at the showcase images of my trained LoRAs:
A friend commented that this looks like superheroes lining up. Exactly right. Although the characters in the Computational Universe have diverse appearances and span stories of varying lengths, they all share a unified theme: they are all heroes. This sounds like a game, which is precisely the effect I want to achieve.
I’m deeply influenced by popular culture. Although I studied philosophy and own many philosophy books, what truly moves me emotionally are the stories of the Bat family. So, the protagonists in my stories, besides being philosophers by nature, are also superheroes by nature.
Some people say I should write stories about ordinary people, but I actually can’t relate to them. At least ordinary people rarely move me – the basic content of my stories is that even superheroes find things very difficult. They don’t just have powerful abilities; they also carry responsibility and sorrow.
AI technology keeps advancing. When studying technology, we must acknowledge that current techniques are likely to become obsolete within a relatively short period; new technologies will need to be adopted. Ideally, there should be dedicated team members discovering new technologies daily, but if you’re a one-person team, you have to work even harder to keep up.
Studying technology is not inferior to studying art technique – I still believe AI is a combination of technology, art, and literature. It’s a comprehensive capability that even requires creators to be polymaths. Only in this way can works with soul and vitality be created.
I spent so long studying character consistency, and then it became obsolete overnight – I have to be philosophical about this. This multi-reference update may be the foundation for unleashing my creativity. Not having to constantly worry about character consistency will free up more energy to focus on narrative details.
You should also consider what kind of story you want to tell once the consistency issues are resolved. Or, if not telling stories, what kind of atmosphere do you want to create? I know where I’m going – I want to see the Computational Universe become a memorable IP that has a lasting spiritual impact.
Character consistency is resolved; IP consistency remains to be developed. This is my research direction: popular culture sci-fi narratives in exotic costumes and alternate worlds. Stories of new-age superheroes exploring the universe. I, indeed, haven’t played fewer games or watched fewer movies.
Follow me and explore AI filmmaking together.
