The Evolution of AI Video Generation: From the Will Smith Spaghetti Disaster to Cinematic Quality

A Technical Milestone Behind a Meme

In the AI video generation community, there's a widely circulated "folk benchmark" — the Will Smith Eating Spaghetti test (Will Smith Benchmark). This seemingly absurd test has become an unofficial yardstick for measuring the capabilities of Text-to-Video models.

Recently, a Reddit user dug up the original "Will Smith eating spaghetti" video generated by ModelScope, paired with that iconic comment: "KEEP THAT AI SPAGHETTI INSIDE MY F*ING MOUTH!" — instantly igniting a wave of nostalgia and technical discussion across the community.

Why did a distorted, mangled video where noodles and mouth become indistinguishable turn into a classic? Because it precisely exposed the core challenges of early video generation models — the breakdown of Temporal Consistency and physical logic. Temporal consistency is one of the most critical challenges in video generation: unlike single image generation, video requires maintaining coherent character appearances, backgrounds, and object states across dozens or even hundreds of consecutive frames. Early models typically generated frames individually or in extremely short windows, lacking cross-frame global constraint mechanisms, causing faces to undergo uncontrollable deformation and flickering between consecutive frames. Physical logic, meanwhile, requires the model to have an implicit understanding of gravity, collisions, soft-body deformation, and other real-world rules — spaghetti as a non-rigid object exhibits extremely complex stretching and drooping behavior when being twirled on a fork and brought to the mouth, far beyond the capability boundaries of early models.

Why the Will Smith Benchmark Matters

A Devilishly Difficult Test Scenario

The action of "a person eating noodles" is practically nightmare-level difficulty for AI. It simultaneously tests multiple core capabilities of video generation models:

Facial consistency: Faces cannot deform or collapse across consecutive frames
Object interaction: The physical relationships between noodles, hands, mouth, and utensils
Dynamic coherence: Smooth transitions through the sequence of chewing and eating motions
Detail stability: Thin, soft objects like noodles are extremely prone to "melting into a blob" during generation

The early ModelScope model failed this test spectacularly — Will Smith's face would twist and distort, noodles merged with his mouth, and the footage was filled with that unmistakable uncanny "AI flavor." ModelScope Text-to-Video was an open-source text-to-video model built on a Diffusion Model architecture, released by Alibaba's DAMO Academy in early 2023. As one of the first text-to-video models made publicly available, its output quality was quite rough by today's standards — low resolution, few frames, severe distortion — but its open-source nature enabled rapid spread across Reddit, Twitter, and other communities, becoming the gateway through which countless users first experienced AI video generation. It was precisely this widespread accessibility that spawned viral content including the "Will Smith eating spaghetti" clip, turning its absurd failures into a beloved classic meme in the AI video generation community.

From Flaws to Fun: Community Culture

Interestingly, the community didn't mock these failures — instead, they embraced them with playful humor. Comments like "SOTA models wish they were half as cool as Modelscope" reflect tech enthusiasts' appreciation for early exploration — those rough outputs are witnesses to the very beginnings of the entire AI video generation field.

The Breakneck Evolution of AI Video Generation

The Massive Leap from Two-Three Years Ago to Today

One user's sentiment captured what many feel: "This was two years ago? Three years ago? I haven't known what time is since the pandemic, but the pace of progress in video generation is truly astonishing."

From the ModelScope era to today, AI video generation has undergone a near-exponential transformation:

Early exploration phase: Early models represented by ModelScope and Runway Gen-1 universally produced outputs with severe distortion, flickering, and logical incoherence
Breakthrough advances: OpenAI's Sora burst onto the scene, demonstrating highly consistent videos up to one minute long, shaking the entire industry. Sora released its technical demonstration in February 2024, with its core breakthrough being the treatment of video as sequences of "spacetime patches," unifying the handling of different resolutions and durations. Through training on extremely large-scale video data, Sora acquired a kind of "emergent understanding" of the physical world — capable of simulating light reflections, object permanence, and other complex phenomena. It also supports native generation at variable durations and resolutions, breaking the fixed-size output limitations of previous models, marking a critical turning point from "toy-grade" to "production-grade" AI video generation.
Flourishing ecosystem: Runway Gen-3, Kling, Luma, Pika, and other products launched in succession, with dramatic improvements in video quality, duration, and controllability

"Staggering" or "Awesome"?

A subtle disagreement emerged in the comments. Some used "awesome" to describe this progress, while others leaned toward "staggering" — implying something shocking, even unsettling.

This word choice precisely reflects the complex sentiment within the industry: on one hand, text-to-video technology brings unprecedented creative possibilities; on the other, such rapid development also creates a dizzying sense of losing control — with deepfakes, content authenticity, and creative ethics issues following close behind. The meaning of Deepfake technology has expanded from its original face-swapping application to encompass any AI-generated media that can pass as real. When video generation models can produce highly realistic footage, the line between real and synthetic grows increasingly blurry. Current industry approaches to this challenge include digital watermarking (such as Google's SynthID), content provenance standards (such as the C2PA coalition's content authentication protocol), and AI detection models. However, the adversarial nature between detection methods and generation technology is fundamentally a continuous arms race, constituting the deepest societal and ethical challenge in this field.

Lessons from "Fail Videos" About AI Video Generation's Development Patterns

Flaws as a Coordinate System for Progress

The Will Smith Benchmark is valuable because it provides a perceptible, intuitive coordinate system for progress. Ordinary people might not understand technical metrics like FID or CLIP Score, but everyone can immediately see that "this time the noodles actually made it into the mouth."

It's worth explaining what these professional metrics mean: FID (Fréchet Inception Distance) evaluates generation quality by comparing the statistical distribution differences between real and generated data in the feature space of deep neural networks — lower values indicate generated results closer to reality. CLIP Score uses OpenAI's CLIP model to measure the semantic alignment between generated content and text descriptions — higher values indicate better adherence to text prompts. However, both metrics have obvious limitations — FID cannot capture temporal coherence, and CLIP Score struggles to assess physical plausibility. This is precisely why intuitive folk tests like "Will Smith eating spaghetti" have irreplaceable value: they make abstract technical progress concrete and tangible, allowing the entire community to track the frontier of video generation models in a lighthearted way.

The Compute and Data Behind Rapid Iteration

The enormous leaps AI video generation achieved in just two to three years are the result of multiple compounding factors:

Fusion of Diffusion Models and Transformer architecture: Particularly DiT (Diffusion Transformer), which has become the mainstream technical approach for video generation. DiT was first proposed by Meta researchers William Peebles and Saining Xie in 2023. Its core idea is replacing the U-Net backbone in traditional diffusion models with a Transformer. Traditional U-Net relies on convolutional operations to capture local features, while Transformer's self-attention mechanism can model global dependencies and naturally extends to the temporal dimension — processing video by dividing it into spatiotemporal patches in a unified manner, enabling models to maintain consistency over longer time spans. OpenAI's Sora is widely believed to employ a DiT-like architecture, which is the key technical foundation for its ability to generate long, coherent videos.
Accumulation and curation of large-scale video datasets, providing models with richer training material
Sustained investment in compute: Though training costs are high, output quality improves exponentially

Conclusion: Cherish Every "AI Eating Spaghetti" Moment

When we marvel today at the cinematic-quality footage generated by Sora and Kling, let's not forget that distorted, hilarious "Will Smith eating spaghetti" video. It's not just a meme — it's a milestone in the history of AI video generation, documenting this field's entire journey from toddling first steps to full-speed sprinting.

As the community puts it, whether you use "awesome" or "staggering" to describe it, this evolutionary journey deserves its place in history. And next time someone runs the Will Smith Benchmark again, we may be surprised to find that the spaghetti has long since made it perfectly into the mouth.