Questioning the capabilities of artificial intelligence proves necessary, especially when it comes to recreating a masterpiece. The attempts to reproduce “The Waterer Watered”, the very first film in history, raise fundamental questions. The stakes go beyond mere technical reproduction; they engage a reflection on the very essence of cinema and artistic creation. Technological advancements, though impressive, fail to capture the quintessence of an iconic work. The result of these efforts reflects a picture of incoherence and disappointment, thus illustrating the current limits of AI in the cinematic field.
An ambitious essay in video generation by artificial intelligence
The ultimate challenge has been set: to reproduce “The Waterer Watered”, the first fictional film. This masterpiece, directed by Louis Lumière in 1895, has been the target of a series of tests by artificial intelligence models, such as Sora from OpenAI, Gen-4 from Runway, Veo-2 from Google, and Kling from Kuaishou. The goal was to determine whether AI could create sequences as smooth and comedic as the original work.
Decomposing the original sequences
To better achieve the objective, the film was segmented into four major sequences. Each scene was meant to capture key moments: the gardener watering his garden, the boy preventing the water from flowing, the impulsive return of the water to the gardener, and finally, the boy’s attempt to be captured by the gardener. Each segment holds comedic potential, a central element of the first filmed fiction.
Initial attempts with Sora from OpenAI
The first approach relied on Sora’s classic text-to-video model. The results, however, proved to be bewildering. The sequences appeared completely out of context, disappointing initial expectations. Seeking a solution, the researchers opted for image-to-video with Sora, providing the models with still images for better coherence. Unfortunately, this method also did not yield satisfactory renders.
AI-assisted colorization
Faced with unsatisfactory results, the experimenters resorted to captures from the original film. These images were then colorized using Google’s Gemini Flash 2.0 Exp, ensuring superior aesthetic fidelity. The colorizations produced vibrant scenes, reminiscent of those that a real shoot could have offered.
Exploration with Gen-4 from Runway
Renewing their efforts, the researchers switched models, turning to Gen-4 from Runway. Using the colorized images as a starting point, they attempted to generate sequences. Despite a slightly more relevant outcome, the attempts were deemed largely below expectations. A second sequence, for example, was so far from the envisioned concept that it sparked disillusionment.
Using Veo-2 from Google DeepMind
Veo-2, the latest from Google DeepMind, was introduced to try to improve the situation. This model allowed for more faithful generalizations, establishing a more virtuous spatio-temporal coherence. The sequences gave a sense of real life, although issues of continuity remained to be lamented, particularly regarding the faces and costumes of the characters.
Outcome and final evaluation
The final phase integrated the review of the images produced by Gemini, seeking to refine them. Each sequence was submitted to the Kling 2.6 model from Kuaishou, hoping for a convergence between aesthetics and respect for the original work. The results proved to be more photorealistic, but unconvincing in terms of narrative. The changes between sequences remained too pronounced, calling into question the continuity of the story.
Despite hours of hard work, the verdict emerged: reproducing films, even the very first one, remains a complex task for contemporary artificial intelligences. These attempts highlight the current limits of video AI, even as its potential seems vast but still embryonic. Researchers, armed with resilience, continue their explorations, hoping to see AI reach new heights.
FAQ on using artificial intelligence to recreate the first historical film
Why has artificial intelligence not succeeded in satisfactorily recreating “The Waterer Watered”?
The advancements of artificial intelligence in video generation are still in the experimental phase. The tested models struggled to capture the narrative and visual coherence of the sequences, leading to a final result far from expectations.
Which artificial intelligence models were used in this essay?
Models such as Sora from OpenAI, Gen-4 from Runway, Veo-2 from Google, and Kling from Kuaishou were tested to attempt to generate the video.
What were the main limitations of the AI models used?
The main issues encountered included incoherence in automated details, such as changes in character appearance and misinterpreted actions, making the sequences difficult to follow.
How did you prepare the sequences for the video generation test?
The sequences from the original film were cut into four main parts, with particular attention given to describing each action, to provide AI models with clear reference points.
What type of images was used to feed the AI models?
Images generated by other AI models and captures from the original film that were colorized were used as a visual base for each sequence to improve the coherence of the results.
What approach proved to be the most promising during the tests?
The method with Veo-2 turned out to be the most promising, managing to generate relatively coherent sequences, although the results still fell significantly short of expectations.
Why is the use of AI to recreate this film considered “cheating”?
Mentions of the use of image captures from the original film and their colorization were utilized, which could be perceived as cheating since it is not an entirely original creation based on artificial intelligence.
What lessons were learned from this artificial intelligence essay?
It was observed that despite progress made, generative video AI is not yet ready to reproduce historical films with the necessary accuracy, highlighting the need for further developments in the field.
Is it possible to expect better results with the evolution of video AI?
Yes, it is highly likely that artificial intelligence models will continue to improve, potentially offering better video creation capabilities in the future.
