Will Smith Eating Spaghetti test
Informal benchmark for text-to-video models
From Wikipedia, the free encyclopedia
The Will Smith Eating Spaghetti Benchmark is an informal benchmark within the artificial intelligence community, used to assess the capabilities of generative video models in rendering realistic human actions and facial expressions. Originating from a widely shared AI-generated video in 2023, which depicted an unnaturally animated render of actor Will Smith eating spaghetti, the test has since been used as an informal reference point to demonstrate the capabilities and limitations of AI-generated video content.[1][2]
Origin
On March 23, 2023, Reddit user chaindrop shared an AI-generated video titled "Will Smith eating spaghetti" on the StableDiffusion subreddit, created using ModelScope's text-to-video tool. The clip depicted a distorted and surreal version of actor Will Smith eating spaghetti.[3] The video quickly gained attention for its uncanny visuals, with Smith's facial features morphing unpredictably and his eating motions appearing unnatural. The unsettling nature of the video led to widespread sharing and discussions across social media platforms.[4][5][6]
Spread
Use as a benchmark
The prompt became a benchmark for evaluating the realism and coherence of AI-generated videos.[9][6] As large language models advanced, developers and other users use the test to identify improvements. In May 2025, Google's Veo 3, an AI video generation tool, produced a rendition of the test that showcased significant enhancements in facial accuracy, motion fluidity, and audio synchronization.[10] Despite the improvements, viewers observed minor inconsistencies, such as unusually crunchy sound effects accompanying the spaghetti consumption.[11][6] A version created with Veo 3.1 in October 2025 was noted by Business Insider for being more realistic.[12] In February 2026, Forbes stated that Seedance 2.0 had passed the test through its response to the prompt.[13]