What Happened
AI video generation startup Higgsfield has reportedly paid a bartender from New Jersey over $1 million for the rights to use his likeness. According to a post by AI influencer Hasaan Toor, the individual received the payment for providing a full-face 3D scan, with no requirement for acting experience, auditions, an agent, or set filming days.
The deal is specifically for training data. Higgsfield is using the bartender's facial biometrics to train its flagship text-to-video model, Diffuse. The model is designed to generate photorealistic video content featuring consistent human characters from textual descriptions.
Context: The Race for Realistic Digital Humans
Higgsfield, founded by former Snap AI researchers, is competing in the rapidly advancing field of generative video. While models like OpenAI's Sora, Runway's Gen-2, and Pika Labs have demonstrated impressive scene generation, a key technical hurdle remains character consistency—maintaining a believable, stable human identity across different shots and scenes.
Acquiring high-quality, legally licensed 3D facial scans is one approach to solving this. By training on a detailed, consistent dataset of one individual's face from multiple angles and under varied lighting, models can learn to generate that specific person more reliably. The seven-figure price tag indicates the premium Higgsfield places on obtaining a clean, versatile, and exclusive dataset for this purpose.
This follows a broader trend of AI companies seeking licensing deals with individuals for voice, likeness, or movement data, moving beyond scraping publicly available information.
The Model: Diffuse
Higgsfield's Diffuse model is not yet publicly available. Based on the company's previous research and statements, it is a diffusion-based model for text-to-video generation. The core technical challenge it aims to address is the "identity preservation" problem in generated video—keeping a synthetic character looking like the same person throughout a sequence.
The use of a high-fidelity 3D scan suggests the training pipeline may involve constructing a detailed neural radiance field (NeRF) or similar 3D representation of the subject. This volumetric data can then be used to synthesize the face from novel viewpoints and under different conditions within the generated videos, providing a strong prior for consistency.
Implications for Data Sourcing
The transaction signals a shift in how AI companies may acquire training data for sensitive biometric domains.
- From Scraping to Licensing: It represents a move toward formal, compensated licensing agreements for personal biometric data, potentially setting a precedent for valuation.
- Market Creation: It could catalyze a new market for "AI model training likenesses," separate from traditional acting or influencer careers.
- Legal and Ethical Frameworks: High-profile deals like this will pressure the industry to develop clearer standards for consent, compensation, and usage rights for digital likenesses used in AI training.
&w=3840&q=75)
