The video introduces the release of Veo 3.1, the latest version following Veo 3 and Sora 2, focusing on testing and comparing these AI video generation models using the Higgsfield AI platform. The presenter highlights that Veo 3.1 is currently in early access, with the full release promising video durations over 30 seconds and 1080p resolution. Using Higgsfield, users can easily generate videos by uploading an image, entering a prompt, selecting the model (V3.1), quality, and aspect ratio, making the process straightforward and cost-efficient. The platform also offers various presets tailored for different social media formats, enhancing usability.
Several examples demonstrate Veo 3.1’s capabilities and limitations. One amusing test involved generating a video of the presenter, which resulted in some quirky inaccuracies like bald hair and exaggerated features, yet still looked fairly convincing overall. Another test involved a digger and a bodybuilder performing complex movements like a front double bicep pose and a backflip. While the front pose was well-rendered, the backflip animation was unsuccessful, showing the model’s current struggles with complex physics and motion consistency. Despite some flaws, the model showed promise in maintaining detail across different movements.
The video also showcases Veo 3.1’s improved audio synchronization compared to previous versions, with an example of the presenter eating wood accompanied by satisfying and relevant sound effects. However, some audio outputs remain nonsensical, as demonstrated by a clip featuring the presenter’s partner where the speech was unintelligible. Additionally, the AI can generate music, exemplified by a catchy song about a Greg’s sausage roll lorry created via ChatGPT, hinting at exciting creative possibilities if longer videos become available.
Comparisons between Veo 3.1, Veo 3, and Sora 2 reveal that while Veo 3.1 is not a massive leap forward, it shows noticeable improvements in video quality, facial detail, and realism. For instance, a video of a man receiving a million-subscriber plaque from a robot was more polished in Veo 3.1, avoiding some of the hand duplication errors seen in Sora 2. However, some animations still appeared less realistic or sensationalized, such as characters’ clothing disappearing or odd uniform choices. Overall, Veo 3.1 is progressing but still has room for refinement.
In conclusion, the presenter expresses mixed feelings about the rapid advancements in AI video generation, acknowledging both the impressive technological progress and the somewhat dystopian implications. They encourage viewers to try out Higgsfield AI to experience the latest models firsthand and invite comments on opinions about AI’s evolving role in creative media. The video serves as both a demonstration and a thoughtful reflection on the current state and future potential of AI-generated video content.
