A recent publication by the InstantX team in Beijing introduces a novel AI image generation method named InstantID. This technology boasts the capability to swiftly identify individuals and generate new images based on a single reference image.
Despite being hailed as a "new state-of-the-art" by Reuven Cohen, an enterprise AI consultant, concerns arise regarding its potential misuse for creating deepfake content, including audio, images, and videos, especially as the 2024 election approaches.
Cohen highlights the downside of InstantID, emphasizing its ease of use and ability to produce convincing deepfakes without the need for extensive training or fine-tuning. According to him, the tool's efficiency in generating identity-preserving content could lead to a surge in highly realistic deepfakes, requiring minimal GPU and CPU resources.
In comparison to the prevalent LoRA models, InstantID surpasses them in identifiable AI image generation. Cohen, in a LinkedIn post, bids farewell to LoRA, dubbing InstantID as "deep fakes on steroids."
The team's paper, titled "InstantID: Zero-shot Identity-Preserving Generation in Seconds," asserts that InstantID outperforms techniques like LoRA by offering a 'plug and play module' capable of handling image personalization with just a single facial reference image, ensuring high fidelity without the drawbacks of storage demands and lengthy fine-tuning processes.
Cohen elucidates that InstantID specializes in zero-shot identity-preserving generation, distinguishing itself from LoRA and its extension QLoRA. While LoRA and QLoRA focus on fine-tuning models, InstantID prioritizes generating outputs that maintain the identity characteristics of the input data efficiently and rapidly.
The simplicity of creating AI deepfakes is underscored by InstantID's primary functionality, which centers on preserving identity aspects in the generated content. Cohen warns that the tool makes it exceedingly easy to engineer deepfakes, requiring only a single click to deploy on platforms like Hugging Face or replicate.
