The creator used Flux 3 within the Hermes Agent app to produce a cohesive short film featuring a consistent character traveling through various distinct worlds, showcasing the model’s ability to generate diverse visual styles from text prompts and images. Despite challenges with character movement and scene transitions, the project was completed in half a day, demonstrating Flux 3’s strengths and limitations in AI-driven video generation.
In this video, the creator explores making a short film using Flux 3, a new multimodal video generation model from Black Forest Labs, integrated within the Hermes Agent desktop app. Flux 3 supports image, video, audio, and action prediction, allowing users to generate video content from text prompts and images. The creator’s goal was to produce a cohesive short film featuring a consistent character traveling through various distinct worlds, testing Flux 3’s ability to maintain character consistency across different artistic styles and environments.
The process began by designing the main character based on the creator’s profile picture, resulting in a simple, marker-style mascot named Toby. The film’s narrative involves Toby moving through different worlds connected by golden portals, including a marker-style world, a realistic New York City street, an anime samurai world, a photoreal volcanic coastline, and a claymation dinosaur world. The creator used key frames to plan each scene and experimented with clip lengths, eventually opting for longer continuous shots to improve seamlessness. While some scenes, like the New York and volcanic beach sequences, turned out well, others required multiple attempts to fix issues such as unnatural transitions and character movement.
One of the more challenging sequences involved a dinosaur world transitioning into a 1910s railway station scene. The Flux 3 model struggled with the dinosaur’s movement, causing visual inconsistencies like the dinosaur’s head turning into a tail. To address this, the creator simplified the animation by having the dinosaur remain stationary while the character interacts with the environment. Despite some connection errors and the need to stitch together multiple clips to maintain narrative flow, the creator managed to produce a visually interesting sequence with period-appropriate reactions and silent cinema style aesthetics.
Further experimentation included a highly stylized stained glass fantasy cathedral world and a realistic ancient Rome setting, showcasing Flux 3’s versatility in handling different visual styles. The film concludes with the character returning to the original marker world, entering a workshop filled with framed images from his adventures, accompanied by a motivational tagline: “If you can imagine it, you can explore it.” The creator emphasized the importance of using reference images to guide the model and shared insights into the challenges of maintaining visual consistency and smooth transitions throughout the film.
Overall, the project took about half a day to complete and involved extensive trial and error, but the creator expressed satisfaction with the final two-and-a-half-minute film. They highlighted Flux 3’s strengths in generating diverse and visually appealing scenes, while also acknowledging the difficulties in video generation, such as clipping, character consistency, and rendering errors. The video serves as both a demonstration of Flux 3’s capabilities within Hermes Agent and a practical guide for others interested in AI-driven video creation.