Veo 3 Crowd Simulation: Prompts, Physics & VFX Pipelines

1. Executive Strategy: Positioning VEO3 in the Generative Landscape

The advent of Google’s Veo 3 represents a watershed moment in the trajectory of generative media, marking a decisive shift from the stochastic, dream-like sequences of early diffusion models to controllable, physically plausible cinematic outputs. Within this broader evolution, the specific domain of Crowd Simulation stands as one of the most rigorous stress tests for any generative system. Unlike the generation of a single subject, where minor inconsistencies can be masked by artistic style, crowd simulation requires the simultaneous management of spatiotemporal coherence for dozens, if not hundreds, of independent agents. This report provides an exhaustive analysis of Veo 3’s capabilities in this high-complexity domain, tailored for technical directors, VFX supervisors, and AI strategists.

1.1 The SEO and Content Ecosystem for AI Crowd Simulation

To understand the impact of Veo 3, one must first analyze the informational landscape it occupies. The current discourse is polarized between hyper-technical research papers and superficial consumer demonstrations. There exists a critical vacuum for "bridge literature"—content that translates the architectural innovations of Latent Diffusion Transformers (DiTs) into actionable workflows for professional production.

The "High-Intent" Search Landscape

Analysis of search behaviors reveals a maturing user base. Early queries focused on generic terms like "AI video generator." Today, high-intent traffic is driven by specific technical pain points. Users are searching for "Veo 3 object permanence settings," "fixing temporal drift in AI crowds," and "integrating Veo 3 with Nuke compositing." This shift indicates a transition from experimentation to integration. Content strategies must therefore pivot from "showcasing" to "solving."

Strategic Content Pillars

To dominate this niche, the content strategy must prioritize utility over novelty.

1. The Physics of Hallucination: Content that explains why artifacts occur (e.g., the "Piano Movers Problem" where bodies deform to fit through gaps) establishes deep authority. It moves the conversation from "the AI made a mistake" to "the model prioritized flow over rigid body dynamics."

2. The "Three-Subject" Threshold: Acknowledging and offering workarounds for the model's limitation in handling more than three primary focal points is crucial for professional credibility.

3. Prompt Engineering as Directing: Framing prompt construction not as captioning, but as technical direction (using JSON structures or specific "camera syntax") appeals to the professional filmmaker.

1.2 The User Psychology of Crowd Generation

The desire to generate crowds stems from a resource constraint. In traditional film production, "extras" are expensive—they require casting, costuming, feeding, and directing. In traditional CGI (e.g., Massive, Golaem), crowds are computationally expensive and technically demanding to set up. Veo 3 promises to bypass both barriers, offering "infinite extras" for the cost of inference. The content strategy must address this value proposition: Veo 3 is not just a creative tool; it is a budgetary compression algorithm. It turns a line item on a budget into a text prompt.

However, this comes with a "Control Tax." The user gains infinite scale but loses precise control. The content strategy must therefore focus on "Steerability"—how to regain control through "Flow" tools, image-to-video anchors, and iterative prompting. By addressing the trade-off between scale and control, we align with the strategic concerns of production heads and studio executives.

2. Technological Evolution: From U-Nets to Spatiotemporal Transformers

The capabilities of Veo 3 in crowd simulation are not accidental; they are the result of a specific architectural lineage that diverges from its predecessors. Understanding this evolution is essential for predicting the model's behavior and limitations.

2.1 The Limitations of 2D Convolutional Approaches

Early video generation models typically relied on 3D U-Net architectures, which were essentially 2D image diffusion models extended into the temporal dimension. While effective for short, simple loops, these architectures struggled with global coherence. In a crowd scene, a U-Net might generate a convincing texture of a crowd, but as the camera panned, individuals would "morph" into one another or disappear entirely. This is because the model's "receptive field"—the area of the video it can "see" and reason about at once—was limited. It could ensure that pixel A looked like pixel B, but it struggled to understand that "Person A on the left" is the same entity as "Person A on the right" three seconds later.

2.2 The Rise of the Latent Diffusion Transformer (DiT)

Veo 3 abandons the U-Net in favor of a Latent Diffusion Transformer (DiT) architecture. This is the same architectural shift that powered the jump from GPT-2 to GPT-3 in language, but applied to visual data.

• Spacetime Patches: Instead of processing pixels, Veo 3 compresses video into "latent representations" (highly compressed mathematical summaries of visual information). These latents are then tokenized into "spacetime patches".

• The Attention Mechanism: The critical advantage of the Transformer is the Self-Attention Mechanism. In the context of a crowd, this allows the model to "attend" to specific features across the entire video sequence simultaneously. The model can mathematically link the "red hat" token in Frame 1 to the "red hat" token in Frame 96, even if the camera has moved or other objects have briefly occluded it. This global attention is the fundamental enabler of Object Permanence in dense scenes.

2.3 Compressed Latent Space and Native Audio

Another evolutionary leap is the unified compression of modalities. Veo 3 encodes both video and audio into a shared latent space. This means the model does not generate video and then "add" sound; it "dreams" the sound and the image simultaneously.

• Synesthetic Generation: When simulating a crowd, the roar of the voices is generated from the same underlying "concept" as the visual of the open mouths. This results in Native Synchronization—the audio swells as the visual density increases, not because of a post-process script, but because the latent representation of "dense crowd" inherently contains both visual and auditory data.

2.4 The "Chain-of-Frames" Reasoning Engine

Perhaps the most significant theoretical contribution of Veo 3 is the "Chain-of-Frames" (CoF) approach. This concept posits that video generation is not just a pixel prediction task, but a reasoning task.

• Visual Logic: Just as an LLM uses "Chain-of-Thought" to solve a math problem step-by-step, Veo 3 uses CoF to solve visual problems frame-by-frame. In a crowd scenario, if Person A is walking towards Person B, the model must "reason" about their future states. Will they collide? Will they step aside?

• Sequential Problem Solving: The CoF mechanism encourages the model to predict the next plausible state based on the trajectory of the previous states. This mimics a physics engine's time-step calculation, but it is probabilistic rather than deterministic. It allows Veo 3 to navigate complex interactions—like a crowd parting for a vehicle—that would baffle simpler models.

3. Core Architecture & The Physics of Simulation

Veo 3 creates a compelling illusion of reality, but it is fundamentally different from the simulation tools used in traditional VFX. Understanding this distinction is vital for managing expectations and workflows.

3.1 Emergent Physics vs. Explicit Simulation

In software like Houdini or Massive, crowd simulation is Explicit. The user defines the physical properties of the agents (mass, velocity, collision radius) and the forces acting upon them (gravity, friction). The software then solves these equations to produce the animation.

In Veo 3, the physics are Emergent. The model has never been "taught" Newton's laws of motion. Instead, it has observed billions of hours of video and learned that "when a foot hits the ground, it stops," or "when two solid objects meet, they do not merge."

• The "Piano Movers" Problem: Research into Veo 3’s capabilities highlights the "Piano Movers Problem"—the challenge of moving a large object through a tight space. An explicit simulator would get the object stuck if it didn't fit. Veo 3, prioritizing the continuation of the video, often chooses to deform the object (squashing it like rubber) or phase it through the wall (clipping) to keep the action moving. This reveals that the model's "World Model" is soft and pliable, prioritizing narrative flow over rigid physical integrity.

3.2 Spatiotemporal Attention: The Engine of Object Permanence

Object permanence—the ability of an object to exist when not visible—is the primary benchmark for crowd simulation quality. Veo 3 relies on Cross-Frame Attention to maintain this.

• Memory Banks: The transformer architecture utilizes "memory banks" to store feature representations of objects. If a character in a crowd walks behind a bus, their "visual embedding" is retained in memory. When the bus passes, the model queries this memory to reconstruct the character.

• The Decay of Identity: However, this memory is not perfect. Benchmarks such as Object Instance Segmentation (OIS) show that tracking accuracy degrades over time. In a 4-second clip, identity preservation is high. In an 8-second clip, "identity drift" sets in—a character might emerge from behind the bus wearing a different colored shirt or with a slightly different face. This is the "Fade Factor" of the latent attention span.

3.3 Collision Detection and The "Clipping" Phenomenon

In gaming terms, "clipping" occurs when the collision mesh of one object fails to interact with another. In Veo 3, clipping is a hallucination caused by the model's uncertainty about depth and solidity.

• Soft Body Bias: In dense crowds, Veo 3 often treats bodies as semi-permeable fields. You will frequently observe hands passing through torsos or feet sliding through the pavement. This is because the model is optimizing for the visual appearance of a crowd texture, not the physical reality of individual agents.

• Comparative Physics: When compared to competitors like Kling 3.0, Veo 3 often shows "softer" physics. Kling 3.0 is noted for a "physics-first" approach that results in harder collisions and better weight distribution. Veo 3, by contrast, favors "cinematic smoothing," which makes for beautiful camera moves but can result in "floaty" or "ghostly" interactions in dense packs.

3.4 The Resolution of Latent Space

The fidelity of a crowd is also limited by the resolution of the latent space.

• Detail Compression: The model compresses the entire video into a smaller latent representation. High-frequency details—like the specific features of a face in the background—are often lost in this compression. When the video is decoded back to pixels, the model must "guess" these details.

• The "Impressionist" Crowd: This leads to the phenomenon where background crowds look convincing at a glance (impressionistic) but terrifying upon close inspection (smudged features, missing limbs). The model allocates its "bandwidth" to the primary subjects, leaving the background as a lower-fidelity approximation.

4. Deep Dive: Crowd Simulation Dynamics

Simulating a crowd in Veo 3 is an exercise in managing chaos. The model's behavior changes drastically depending on the density, diversity, and activity level of the requested crowd.

4.1 The "Three-Subject" Threshold

Extensive testing and user reports have identified a recurring limitation in Veo 3: the "Three-Subject Limit." The model is architecturally optimized to handle 1-3 primary "Subject" tokens with high fidelity.

• The Hierarchy of Attention: When a prompt requests "a dense crowd," the model essentially creates a hierarchy:

  1. Tier 1 (Focus): 1-3 characters who receive full attention, maintaining consistent faces, clothing, and physics.

  2. Tier 2 (Mid-Ground): A ring of characters with consistent silhouettes but generic or morphing facial features.

  3. Tier 3 (Background): A "texture" of humanity—blobs of color and motion that mimic the idea of a crowd but lack individual coherence.

• Production Implication: If a shot requires five distinct characters interacting, Veo 3 will likely fail to keep them all consistent. It will merge characters or lose track of one. Professional workflows bypass this by generating the crowd as a background plate and compositing specific "hero" characters (generated separately on green screen) into the foreground.

4.2 Handling Diversity and The "Clone Effect"

Generative models suffer from a phenomenon known as Mode Collapse, where the model converges on a single "safe" output. In crowd scenes, this manifests as the "Clone Effect."

• Homogeneity: A prompt for "a crowd of business people" often results in twenty variations of the same man in a grey suit. This is because the model's statistical average for "business person" is narrow.

• Prompting for Heterogeneity: To combat this, users must explicitly prompt for diversity. Instead of "a diverse crowd," the prompt must be granular: "A chaotic mix of ages, ethnicities, and fashion styles; some in suits, some in casual wear, varying heights and builds." This forces the model to sample from wider areas of its latent space, breaking the visual monotony.

• Bias Filters: Veo 3 includes safety filters to ensure demographic representation, but these can sometimes feel forced or artificial if not guided by the prompt. Users have noted that the model sometimes over-corrects, creating an implausibly diverse crowd for a specific historical or geographic setting unless constrained by specific keywords.

4.3 Crowd Flow and Vector Control

Controlling the directionality of a crowd is one of the hardest tasks in Veo 3.

• Vector Chaos: Without specific instruction, a generated crowd tends to "mill about"—small, random movements that sum to zero net motion. This looks like a cocktail party but fails for a transit hub or a march.

• Directional Verbs: To achieve flow, the prompt must use strong directional verbs that imply vectors. "Marching," "Fleeing," "Streaming," "Rushing." These words align the motion vectors of the latent patches.

• The "River" Analogy: It is helpful to think of the crowd as a fluid. Prompts like "Flowing like a river around the obstacle" often yield better results than "Walking around the obstacle," because they tap into the model's understanding of fluid dynamics, which it often applies to crowd movement.

4.4 Occlusion and Re-emergence

Occlusion—when one object blocks another—is the ultimate test of the "Chain-of-Frames" reasoning.

• The Tunnel Test: If a crowd walks into a tunnel and comes out the other side, Veo 3 often generates new people emerging rather than the same people. The memory of the specific individuals is lost in the darkness of the tunnel (the latent representation becomes too noisy).

• Improving Permanence: Shortening the duration of occlusion helps. If the occlusion lasts less than a second (e.g., passing behind a pillar), the attention mechanism can bridge the gap. If it lasts seconds, the link is broken.

5. Prompt Engineering: The Director's Code

To navigate the complex architecture of Veo 3, the user must evolve from a "Prompter" to a "Technical Director." The prompt is not just a description; it is the code that programs the latent space.

5.1 Layered Prompting: The "Director's Brief"

Successful crowd simulation requires a Layered Prompting strategy. This involves breaking the prompt into distinct functional blocks, ensuring the model attends to each aspect of the scene.

Layer 1: The Anchor (Subject/Focus)

Even in a crowd shot, the model needs a focal point to ground its composition.

• Prompt: "Focus on a weary nurse in blue scrubs, standing still..."

Layer 2: The Environment (Context)

This defines the "container" for the crowd.

• Prompt: "...in the center of a bustling, neon-lit Shibuya crossing at night. Rain slicked pavement reflecting the lights."

Layer 3: The Dynamics (Action/Flow)

This defines the energy and vector of the crowd.

• Prompt: "...surrounded by a chaotic sea of commuters rushing in all directions. Fast-paced motion, blurring past the camera. The crowd creates a vortex around the nurse."

Layer 4: The Lens (Style/Technical)

This is crucial for hiding artifacts.

• Prompt: "Cinematic lighting, heavy motion blur, shallow depth of field, 85mm lens, bokeh in background."

  • Insight: Motion Blur is the crowd simulator's best friend. It masks the low-fidelity faces of the background agents, making the shot look realistic rather than uncanny.

5.2 The JSON Hack: Structuring for Precision

Advanced users have discovered that structuring prompts in a pseudo-code format (like JSON) can significantly improve the model's adherence to complex instructions. While the model doesn't "execute" the code, the structure forces the text encoder to treat each element as distinct, preventing "concept bleed" (e.g., the "blue" of the nurse's scrubs bleeding into the crowd's clothing).

5.3 Camera Syntax and Directional Keywords

The camera is the viewer's avatar in the latent space. Controlling it is essential for selling the scale of a crowd.

• "Parallax": Using the keyword "parallax" forces the model to separate the foreground layers from the background layers. In a crowd shot, this creates the 3D depth effect where foreground people move faster than background people.

• "Rack Focus": This command tells the model to shift the focus plane. "Rack focus from the shouting man in foreground to the silent crowd behind." This is a powerful narrative tool that Veo 3 interprets surprisingly well.

• "Tracking" vs. "Orbiting":

  • Tracking: "Tracking shot following the crowd." This is safe. The camera moves with the vectors, minimizing relative motion and thus minimizing artifacts.

  • Orbiting: "Camera orbits around the stationary crowd." This is high-risk. The model must regenerate the crowd from 360 degrees, which often leads to massive object permanence failures (people disappearing as the camera rotates).

5.4 The "Character Bible" & Ingredients

To maintain consistency of specific characters within a crowd across multiple shots, Veo 3.1’s "Ingredients" feature (part of Google Flow) is essential.

• Canonical Reference: Users upload a "Canonical Image" of their hero character. This image is used as a conditioning signal (like an IP-Adapter in Stable Diffusion).

• The "Bible" Strategy: For text-only consistency, create a "Character Bible"—a block of descriptive text that is pasted verbatim into every prompt. "Man, 30s, scar on left cheek, wearing tattered red hoodie." Any deviation in the wording can cause the model to drift.

6. Native Audio & Soundscapes in Crowds

Veo 3’s ability to generate Native Audio fundamentally changes the crowd simulation workflow. It moves audio from "Post-Production" to "Pre-Visualization."

6.1 The "Cocktail Party Effect" and Native Mixing

The "Cocktail Party Effect" refers to the human brain's ability to focus on a single voice in a noisy room. Veo 3 simulates this through its joint audio-video attention.

• Contextual Volume: When the camera is close to a subject in a crowd, the model automatically mixes their voice louder than the ambient noise. When the camera pulls back to a wide shot, the individual voice is subsumed into the "roar" of the crowd. This is not a programmed foley effect; it is an emergent property of the model learning from movie scenes where this audio dynamic is present.

• Automatic Foley: The model generates "incidental" sounds automatically. If the crowd is on gravel, the footsteps sound crunchy. If on pavement, they sound sharp. This "automatic foley" saves hours of sound design time for background ambiance.

6.2 Spatial Audio Limitations

However, Veo 3’s audio has limitations.

• Screen-Centricity: The audio is generally "screen-centric"—it reflects what is visible in the frame. It does not truly simulate 360-degree spatial audio (Ambisonics). If a car is approaching from behind the camera (off-screen), the model usually won't generate the sound until the car enters the frame, because the visual tokens for the car don't exist yet.

• The "Ventriloquist Effect": In dense crowds, precise lip-sync is a major challenge. The model might generate a voice line but animate the wrong mouth, or animate three mouths for one voice.

  • Mitigation: For dialogue in crowds, use Over-the-Shoulder shots where the speaker's mouth is hidden, or use the "Colon Syntax" (Man in red says: "Hello") to explicitly bind the dialogue to a specific subject description.

6.3 Environmental Audio Prompting

To get high-quality crowd audio, prompts must be as layered as the visuals.

• The "Layered Soundscape" Formula:

  • Bad: "Audio: Crowd noise." -> Result: Generic white noise wash.

  • Good: "Audio: Distinct muffled conversations, occasional laughter, clinking of glasses, distant traffic hum, footsteps on hardwood floor, specific cough close to microphone."

  • Mechanism: Specific keywords like "glass clink" or "cough" trigger distinct audio tokens that puncture the noise floor, creating a sense of depth and reality.

7. Workflows & Pipelines: Integration for Professionals

For Veo 3 to be useful in a professional environment, it must integrate into existing pipelines. It is rarely a "text-to-final-pixel" solution.

7.1 The "Flow" Ecosystem: Pre-Visualization and Assembly

Google Flow acts as the bridge between the raw model and the edit timeline.

• Scenebuilder: This tool allows users to string together multiple generated clips to build a sequence. For crowd scenes, this is vital. A standard 8-second generation is often too short to establish the mood of a large event. Scenebuilder allows the user to extend the clip, maintaining the flow of the crowd.

• The "Freeze" Technique: A common workflow for controlling crowds is to use an Image-to-Video workflow.

  1. Generate a high-resolution image of a crowd (using Midjourney or Imagen 3) where the composition is perfect.

  2. Import this image into Veo 3/Flow as the starting frame.

  3. Prompt: "Crowd comes to life, subtle motion, breathing, shifting weight."

  • Benefit: This guarantees the visual fidelity of the faces (which image models are better at) while using Veo 3 for the motion. It prevents the "melting face" artifacts that often occur when Veo 3 generates the crowd from scratch.

7.2 Post-Production: Upscaling and Compositing

Raw Veo 3 output (often 1080p, compressed) is not broadcast ready.

• Upscaling: Tools like Topaz Video AI are standard. For crowds, the "Proteus" or "Iris" models in Topaz are recommended as they are trained to recover facial details. This step can often "fix" the blurry faces in the background of a Veo 3 shot.

• Frame Interpolation: Veo 3 typically outputs at 24fps. For slow-motion crowd shots, optical flow interpolation (using Twixtor or DaVinci Resolve’s Speed Warp) is necessary.

• The "Pass" Problem: Veo 3 does not output render passes (Depth, Alpha, Normal, Motion Vector). This makes compositing difficult.

  • The "Depth Estimation" Workaround: Compositors use AI depth estimation tools (like Depth Anything or built-in Nuke AI nodes) to generate a Z-depth map from the Veo 3 video. This allows them to insert atmospheric fog into the crowd or place 3D text behind the foreground characters, faking a 3D composite.

7.3 Hybrid Workflows: Veo as Texture

A growing trend in high-end VFX is using Veo 3 as a "Texture Generator" for traditional simulations.

• The Workflow:

  1. Create a low-fidelity crowd simulation in Houdini or Blender using blocky avatars. This ensures perfect physics and collision.

  2. Use ControlNet (or Veo 3’s "Structure" guidance, if available via API) to "paint" the realistic crowd over the blocky simulation.

  3. This combines the rigorous physics of explicit simulation with the photorealism of generative AI.

8. Pitfalls, Hallucinations, and Troubleshooting

Understanding the failure modes of Veo 3 is as important as understanding its capabilities.

8.1 The "Melting Face" Artifact

In the background of crowd shots, faces often devolve into impressionistic smears or "melt."

• Cause: The latent resolution is insufficient to encode high-frequency facial details for small objects.

• Fix:

  1. Depth of Field: Aggressively prompt for "bokeh" or "shallow depth of field." If the background is blurry, the lack of detail is a feature, not a bug.

  2. Grain: Adding film grain in post-production helps mask the "plastic" look of the melted faces.

8.2 The "Limb Spawning" Glitch

Crowds often feature people with extra arms or legs, or legs that merge into each other.

• Cause: The model struggles to separate the "leg" tokens of overlapping people.

• Fix: Use Negative Prompts: "extra limbs, malformed anatomy, morphing, fusing bodies, bad hands, polydactyly."

8.3 Temporal Drift (The "Shifting Shirt")

A character might enter the frame wearing a red shirt and leave wearing a maroon one.

• Cause: The memory bank's attention decays over time.

• Fix: Keep shots short (<5 seconds). Use the "Ingredients" feature to lock the character's appearance. Avoid lighting changes (e.g., strobe lights) that confuse the model's color consistency.

8.4 Audio Hallucinations

The model sometimes generates sounds that aren't there.

• Ghost Voices: Hearing speech when no one is talking.

• Studio Laughter: Hearing a "sitcom laugh track" if the scene feels like a comedy.

• Fix: Explicit Negative Audio Prompts: "No dialogue, no speech, no laughter, no music, ambient noise only."

9. Comparative Analysis: Veo 3 vs. The Field

9.1 Veo 3 vs. Traditional VFX (Houdini / Massive)

Insight: Veo 3 is not a replacement for Massive in scenes like the Lord of the Rings battles where specific soldiers need to fight specific orcs. It is a replacement for Stock Footage—generic crowds, city streets, and background ambiance.

9.2 Veo 3 vs. Sora 2 vs. Kling 3.0

• Google Veo 3:

  • Pros: Best-in-class integration with Google ecosystem (YouTube data), native audio is superior, "Flow" offers better editorial control.

  • Cons: Soft physics, prone to "gumminess" in dense crowds.

• OpenAI Sora 2:

  • Pros: Superior narrative coherence, often better at "dream-like" transitions, strong temporal consistency.

  • Cons: Slower generation, restricted access, often struggles with realistic "gritty" textures compared to Veo.

• Kling 3.0:

  • Pros: "Physics-First" Engine. Kling is widely regarded as having the best collision detection and weight simulation. Its crowds feel "heavier" and grounded.

  • Cons: Audio integration is weaker than Veo. UI is less accessible for Western markets.

10. Future Outlook: The Trajectory of "SkyReels" and World Models

The future of crowd simulation lies in the convergence of Generative Video and World Models. Research into systems like SkyReels-V3 suggests the next generation of models will not just predict pixels, but will build an internal 3D representation of the scene.

• The "Game Engine" Convergence: We are moving toward models that are "Neural Game Engines." They will allow users to "spawn" a crowd (generative) but then control it with the rigidity of a physics engine (simulation).

• Hybrid AI: The immediate future is Hybrid workflows. We will see pipelines where a rough 3D simulation drives the structure, and Veo 3 "skins" the simulation with photorealistic texture.

• Economic Impact: This technology will likely "hollow out" the mid-tier VFX market. Tasks like "roto," "crowd tiling," and "background plate generation" will be automated. The value will shift to Editorial and Creative Direction—the ability to curate and assemble these infinite synthetic assets into a coherent story.

11. Conclusion

Google Veo 3 represents a fundamental democratization of scale. It allows a single creator to summon a cast of thousands with a sentence. However, it is a tool of Impressionism, not Precision. It excels at capturing the feeling of a crowd—the chaotic energy, the motion, the roar—but fails at the rigorous logic of individual agency.

For the professional, Veo 3 is best viewed not as a "Simulation Engine" but as a "Infinite Stock Footage Library" that can be directed. Success requires a mastery of Prompt Engineering (to guide the dream), Technical Knowledge (to hide the artifacts), and Workflow Integration (to polish the output). As the architectures evolve from simple diffusion to physics-aware World Models, the gap between "dreaming" a crowd and "simulating" one will continue to close, ushering in a new era of synthetic filmmaking where the only limit is the director's ability to describe the scene.