- AI-generated videos often become inconsistent over time due to a problem called drift.
- Models trained on perfect data struggle when processing imperfect real-world inputs
- EPFL researchers developed retraining with error recycling to limit progressive deterioration
AI-generated videos often become inconsistent over longer sequences, an issue known as drift.
This problem occurs because each new frame is generated based on the previous one, so small errors such as distorted objects or slightly blurred faces are amplified over time.
Large language models trained only on ideal datasets have a hard time handling incomplete input. So videos usually become unrealistic after a few seconds.
Improve AI performance by recycling errors
Generating videos that maintain logical continuity over time remains a major challenge in this field.
Now, researchers at EPFL’s Visual Intelligence for Transportation (VITA) laboratory have introduced a method called retraining by error recycling.
Unlike traditional approaches that try to avoid errors, this method intentionally feeds the AI’s own mistakes back into the training process.
By doing so, the model learns to correct the error in future frames, limiting the progressive degradation of the image.
This process involves generating the video, identifying discrepancies between the generated frames and the intended frames, and retraining the AI on these discrepancies to improve future output.
Current AI video systems typically produce sequences that remain realistic for less than 30 seconds before the shape, color, and motion logic degrade.
By integrating error recycling, the EPFL team could create videos that withstand drift over long periods of time, potentially removing strict time constraints on generated videos.
This advancement will enable AI systems to create more stable sequences in applications such as simulation, animation, and automated visual storytelling.
This approach addresses drift, but does not eliminate all technical limitations.
Recycling and retraining errors increases computational complexity, so continuous monitoring may be required to prevent overfitting to certain mistakes.
Large-scale deployments can face resource and efficiency constraints, as well as the need to maintain consistency across diverse video content.
Whether feeding AI its own errors is really a good idea remains unclear, as this method can introduce unexpected biases and reduce generalizability in complex scenarios.
Developments at VITA Lab show that AI can learn from its own errors, potentially extending time limits for video generation.
However, it is still unclear how this method performs in external controlled tests and creative applications, and caution should be used before assuming that this method completely solves the drift problem.
via tech explorer
Follow TechRadar on Google News and Add us as a preferred source Get expert news, reviews, and opinions in your feed. Be sure to click the follow button!
Of course you can also do Follow TechRadar on TikTok Check out news, reviews, and unboxings in video format and stay updated regularly. whatsapp Too.
