Google has recently unveiled its new AI video generator, Veo 2, during the hype surrounding the launch of OpenAI’s Sora. While Sora became available for general use only earlier this month, reviewers and early testers are already claiming Google's offering outshines OpenAI's model. This narrative is becoming increasingly compelling as users share side-by-side comparisons of outputs generated by both platforms against common prompts.
One significant factor contributing to Google’s edge appears to be its association with YouTube, which has provided the vast dataset used to train Veo 2. According to TED host and former Google employee Bilawal Sidhu, the AI's ability to interpret prompts appears markedly superior. He cited his experience with the prompt, "Eating soup like they do in Europe, the old fashioned way," which produced far more coherent and visually appealing results on Veo 2 than on Sora, which resulted in unsettling visuals. Sidhu shared his findings on social media: "Veo 2 produced something impressive, whereas Sora generated something more terrifying."
For another comparison, Sidhu uploaded the prompt used by renowned YouTube tech reviewer Marques Brownlee (MKBHD), which involved capturing, "A side scrolling shot of a rhinoceros walking through a dry field of low grass plants." Sidhu noted, "It’s not even close," highlighting Veo's superior adherence to the prompt compared to Sora.
Meanwhile, EasyGen founder Ruben Hassid shared his own experiment where he prompted the models to simulate someone cutting a tomato. The result? While Veo 2 displayed accurate knife movement, showing it seamlessly cutting through the tomato, Sora rendered the knife cutting through the hand instead. Hassid quipped, "Comparing Sora and Veo 2 is like comparing a bike to a starship," illustrating the substantial technological gap perceived by early users.
Despite these comparisons stemming from cherry-picked examples, the consensus among AI enthusiasts seems to glean favor for Google’s video creation tool. Justine Moore, partner at Andreessen Horowitz, shared her insights after several hours of testing both models. According to her, Sora tends to "bias more toward motion," producing over-the-top dynamics, whereas Veo displays heightened accuracy and physical realism. She concluded, "A larger percentage of clips from Veo are usable," requiring less post-processing.
Interestingly, Google has been forthright about employing YouTube data to train Veo but has maintained stringent policies preventing other companies from accessing the same resources. This is contrasted by reports of OpenAI’s utilization of YouTube data to bolster its models, which has been of concern, as highlighted by YouTube’s CEO Neal Mohan who noted, "OpenAI doing this would violate our policies."
With the advancements of AI video generation continuing at breakneck speed, Pika Labs recently unveiled version 2 of its video model, dubbed Pika 2. This latest iteration brings with it not just enhancements to motion and realism but also introduces innovative tools aimed at ease of content creation, making it one of the leading platforms within this burgeoning sector. Early testing indicates Pika 2's significant improvements over its predecessor, maintaining its position among the best alongside models like Kling, MiniMax, Runway, and even Sora.
Ryan Morrison, tech journalist and AI editor for Tom's Guide, has had the chance to experiment with Pika 2. He praised the platform for its unique features, including the ability to add “ingredients” to enrich generated content, alongside pre-configured templates, appealing to users who desire more control.
Morrison’s notable test prompt involved generating visuals of "A dog wearing sunglasses traveling on a train." Surprisingly, Pika fared well, keeping the dog stationary as it portrayed external rapid motion accurately, demonstrating impressive control not seen with Sora or Kling. The test continued with Morrison inserting himself as part of different scenarios, successfully generating videos depicting himself against backdrops of AI-generated landscapes—all without any systemic hiccups or inaccuracies.
Among his more whimsical prompts, he asked for "A cat on the moon wearing a spacesuit with Earthrise in the background." This request created visuals reminiscent of commercial studio ident designs, displaying Pika 2’s elevation over standard models and its capacity for creativity.
After extensive testing across various platforms, Morrison emphasized the leap AI video generation has made. Less than one year ago, the technology struggled to produce coherent 2-second clips. Today, tools like Pika 2 and Google’s Veo 2 yield outputs so refined they could be mistaken for actual footage shot with professional equipment.
It’s clear the race among AI video generation tools is heating up, with Google and Pika Labs moving to the forefront as they redefine the standards of video creation efficiency and realism. While early feedback has heralded Google's Veo 2 as the current standout, Pika 2 shows promising potential to challenge this notion as developers and users alike continue to explore these cutting-edge innovations.