How To Create Games With AI: Happy Oyster World Model

We're at a point in time with AI where soon anything will be able to be generated on demand. Images, videos & 3d models are already generated on demand with lightning fast speed. This has transformed some industries such as advertising, marketing, e-commerce, and more. I also see a world where playable high quality video games will be generated on demand with world models. fal has spent years building the inference layer for generative media. In 2023, we shipped 30 FPS, low-latency, action-controlled diffusion APIs, years before anyone else. Fast forward to today and we've been powering foundational model companies across image, video, real-time speech-to-speech and action-controlled world models. Now we're opening it up. World Model Accelerator is the system behind it all.

You are going to want to read this guide to the end if you want to learn how to create some really cool world model games using fal.

We're still in the early innings of generated output quality with world models, but I see that in the next few years this will rapidly advance. Today, we are launching the happy oyster world model engine creator from Alibaba. This is the first playable world model that we are launching on our platform.

Choose From Featured Worlds Or Make Your Own

In fal, there are some featured worlds, like Sunny Meadow, Neon City, Ancient Forest, Desert Ruins and more. Or you can create your own game in less than 60 seconds. A lot of these are open world simulations where you can walk around and interact with the environment. However, there are no limits, and this is up to your imagination on what you can create.

Making Your Own World Model Adventure Game

Go to compose on fal for the happy oyster model: https://fal.ai/models/alibaba/happy-oyster.

1A. Open World Crime Action Game (Add Scene Reference Image)

This workflow is ideal if you want to generate the concept starting from a specific real-world location or a custom character's face as a reference. If you don't need to generate the game with a specific real-world location or a custom character's face as a reference, then read 1B.

A key tip to make this work is to ask it to use a tight over-the-shoulder camera positioning where the player's back is facing you because we are generating a third-person game where they are walking away from the camera.

Prompt: Use the first reference image as the primary gameplay composition and visual language: a high-budget, cinematic third-person open-world crime game with a low, centered follow camera positioned directly behind the playable subjects. Replace the vehicle with the two adult men from the third reference image, walking side by side away from the camera. Keep both men fully rear-facing at all times, showing only their backs, shoulders, backs of their heads, swinging arms, and legs. Preserve their hairstyles, builds, and clothing colors: one in a brown crewneck sweatshirt, dark jeans, and light sneakers; the other in a blue crewneck sweatshirt, dark pants, and dark sneakers. Remove all readable logos. Use the second reference image as the exact San Francisco location: tram tracks, overhead transit wires, dense historic and modern architecture, streetlights, trees, downtown intersections, and the distant clocktower. Make the environment feel like a living, reactive open-world city. Cars accelerate, brake, turn, and pass close by. Streetcars move along the rails. Cyclists weave through traffic. Pedestrians cross, gesture, enter buildings, and react to the two players. Emphasize gritty photorealistic game graphics, detailed pavement, realistic vehicle reflections, dense traffic simulation, cinematic late-afternoon sunlight, dramatic shadows, subtle atmospheric haze, saturated urban color grading, natural character animation, and a polished AAA open-world gameplay feel. No visible faces, weapons, violence, HUD, subtitles, readable text, or branded signs.

In fal Sandbox, you will have to click Edit multiple images, and then you can add the images below as references. You can also pick your desired image editing models like nano banana 2, GPT image 2, and Seedream 5 Pro.

Run this prompt in fal Sandbox

Image output link

Head over to the playground endpoint for Happy Oyster and add in the world description and the first frame reference image we created. Use the same prompt we used to generate the image to create the world model. Make sure you are in the compose tab and select third person perspective. Hit create world, and then in a couple of minutes you will have a playable world model.

Game ID: g3OkY21JhQ7KUH_uNUPP_V7wUc4a2pQIQmx4Q8OO2yE

You can control the game by using your W, A, S, D controls to walk. You can also use Shift to sprint, Space to jump, F to fight, and your arrows to change the perspective of where you are looking.

Here's the output:

Pro tip: You can generate this whole sequence with Claude code + your fal API key. You can even go back and ask Claude to upscale it or add music.

add some upbeat seed audio to the example-5.mp4 clip I just generated. use my fal API key for this and https://fal.ai/models/bytedance/seed-audio-1.0

1B. Open World Crime Action Game (Text to Image Reference)

My other idea is to create a cinematic open-world crime-action game set in a vibrant, modern coastal city like Miami.

Prompt: A cinematic third-person open-world action game scene set in a fictionalized Miami, Florida in the style of an online crime action game. Use a tight over-the-shoulder camera positioned directly behind the player character, slightly above the right shoulder. The character's upper back, shoulder, head, and part of one arm dominate the foreground, while the busy tropical city stretches ahead. Keep the camera close to the player at all times, clearly establishing a playable third-person perspective rather than a wide cinematic shot. The protagonist walks along a sunlit boulevard lined with pastel Art Deco hotels, towering palm trees, neon storefronts, outdoor cafés, luxury condos, and crowded sidewalks. Sports cars cruise through traffic, motorcycles weave between lanes, pedestrians cross the street, and distant boats move across turquoise water. The character wears modern South Florida streetwear: a loose patterned shirt, fitted pants, clean sneakers, sunglasses, and subtle jewelry. Warm late-afternoon sunlight creates long shadows, glossy reflections, and humid golden haze. Palm leaves sway in the ocean breeze while distant helicopters cross the skyline. Use highly detailed, photorealistic graphics, realistic character movement, dense street activity, cinematic lighting, tropical color grading, and immersive environmental sound. No weapons, violence, readable logos, branded signs, or recognizable characters.

(Make sure to set your aspect ratio to 16:9)

Then you're gonna want to head over to our Happy Oyster playground page and click on compose, and then add in your world description prompt (which was the same as my original image generation prompt), choose a third-person perspective, add in your first frame, and hit create world. Use the same method as 1A.

Game ID: g3OkY21JhQ7KUH_uNUPP_d33IrluF1srJvPzCJEDlPc

2. Create A Cinematic Arcade Racing Game

For this second game I used this prompt.

Prompt: A cinematic third-person arcade racing game scene on a wide, snow-covered mountain highway. Use a low chase camera positioned directly behind a bright orange exotic supercar, keeping the rear wing, aerodynamic vents, glowing taillights, and aggressive diffuser dominant in the foreground. The car speeds forward at extreme velocity, centered between the yellow road lines as its tires spray slush, ice particles, and mist across the reflective pavement. A burst of flame shoots from the exhaust during acceleration.

Ahead, a massive avalanche crashes across the road, filling the horizon with dense snow clouds, flying rocks, shattered ice, and airborne debris. The highway curves gently through an expansive alpine landscape surrounded by snow-covered hills, scattered pine trees, wooden barriers, red safety fencing, and enormous layered rock formations. Emphasize constant movement: drifting snow, spinning debris, vibrating bodywork, motion blur along the road edges, tire spray, and dynamic camera shake caused by speed.

Use polished, high-budget racing-game graphics with realistic vehicle reflections, detailed snow textures, dramatic environmental destruction, crisp winter sunlight, deep blue skies, volumetric clouds, and strong contrast between the orange car and the icy environment. The scene should feel fast, dangerous, responsive, and playable rather than like a static automotive photograph. No readable logos, branded vehicles, interface elements, subtitles, pedestrians, or visible drivers. Outdoor lighting, 8k HD. single car only. bright orange.

Using the same method of creating a custom world above, I created a game called Frozen Velocity via happy oyster.

Game ID: g3OkY21JhQ7KUH_uNUPP_YWQdTL3ubcz_QYS7Ljg27A

3. Create A Cartoon Kart Racing Game

Prompt: A vibrant third-person arcade racing game scene viewed from directly behind a small bear character riding a futuristic sport motorcycle. The rider speeds across a wide, elevated racetrack made of glossy, multicolored glass-like panels in turquoise, yellow, pink, green, and blue. Bright electric boost trails spark from the rear wheel, emphasizing rapid acceleration and energetic motion. The track curves sharply to the left through a fantastical tropical environment filled with palm trees, waterfalls, rocky cliffs, oversized balloons, animated billboards, stadium lights, and playful dinosaur sculptures.

Use a low, centered chase camera positioned close behind the motorcycle, creating an immersive gameplay perspective. Include sweeping metallic track structures, transparent loops, illuminated guardrails, black-and-yellow chevron signs, and distant racing platforms. Add a clean game interface with a simplified white minimap on the right and a small lap counter in the lower-left corner reading "3/3." Keep the interface minimal and fully fictional, with no recognizable branding.

Render the scene in polished 3D animation with highly detailed materials, cinematic lighting, soft motion blur, bright midday sunlight, vivid blue skies, large white clouds, colorful reflections, and a cheerful high-energy atmosphere. Wide 16:9 composition, premium console-game visuals, sharp focus, dynamic speed, playful futuristic racing aesthetic.

Game ID: g3OkY21JhQ7KUH_uNUPP_YWQdTL3ubcz_QYS7Ljg27A

Use the same instructions as above to create a custom game where you add in your single reference image, and then you can generate a really nice world model game.

World Models vs Game Engines

A world model predicts what should happen next in an environment based on the current scene and the user's actions, then generates the next visual frame from that learned understanding. A traditional game engine works differently: it has hand-coded rules, explicit game state, physics systems, assets, and rendering logic, so when you press a button, the engine runs math and draws the next frame according to those rules. A world model is impressive because it learns the behavior of the world from data instead of relying on a manually programmed engine. In the future, this could mean more on-demand, personalized, interactive digital experiences where worlds are generated or adapted around each player.

With happy oyster, personalization and unique game state surface in two different ways, director mode & adventure mode.

Director Mode with Happy Oyster

Director Mode is a mode where there is one scene happening in front of you, and then you can input a prompt and then, the scene will change and shift according to whatever new visuals you would like to see.

Adventure Mode with Happy Oyster

With Adventure Mode, you can get really creative and build your own custom world simulation. In Adventure mode, you have can move forward, back, and sideways and sprint, jump, attack or look in other directions. You can choose to play in a first-person perspective or a third-person perspective. My favorite part of Adventure Mode is the fact that you can upload custom starting reference images that will become the main starting state for your game. You can also add in custom prompts for how you want the game engine to react and work. In this guide, I'm going to take you through some adventure mode experiences and the prompts that I use to build them out.

How Happy Oyster Works

Happy Oyster is impressive because it is not a traditional video game with a fixed map, prebuilt levels, and every interaction programmed in advance. It is a world model that generates a playable environment and reacts to the player's actions in real time.

A normal game works like this:

You press a button → the game engine updates the game state → the graphics engine renders the next frame.

The game relies on programmed systems for movement, physics, collisions, characters, environments, camera behavior, and interactions.

Happy Oyster works more like this:

You press a button → the AI predicts what should happen next → it generates the next part of the experience.

When you move forward, the model generates how the environment should continue. When you turn the camera, it predicts what should appear in that direction. When you jump, sprint, attack, or change the scene with a prompt, the generated world responds.

Think of it like an AI that has learned what interactive worlds are supposed to look and behave like. Instead of a developer manually building every street, forest, racetrack, or character interaction, the model generates the experience from a starting image, a world description, and the player's controls.

For example:

Player presses forward → the world continues ahead.

Player turns left → the view shifts and reveals what is on the left.

Player jumps → the character rises and lands.

Player enters a new prompt → the scene changes around them.

This is what makes Happy Oyster different from a normal text-to-video model. It is not simply creating a finished video that you watch from beginning to end. It is generating a world that you can move through, explore, and direct.

The AI is effectively dreaming an interactive environment around the player. It is not the same as a fully hand-built game with perfectly defined rules and persistent game state. It is a generated simulation that responds to your movement, camera controls, reference image, and prompts.

Conclusion

Happy Oyster offers an early look at what AI-generated games could become. Instead of choosing from a fixed map or playing through a world that developers built in advance, you can start with a prompt and a reference image, then generate an interactive environment that responds as you move through it. The technology is still early, and these worlds do not yet have the consistency, control, or depth of traditional games, but the direction is clear: world models are beginning to turn generative video into something playable.

The most exciting part is how quickly you can experiment. You can create a cinematic open-world city, an arcade racing game, a colorful kart racer, or something completely original without building a traditional game engine from scratch. Happy Oyster is available on fal, where you can create your own world, test different prompts and reference images, and experience what real-time generative gameplay already looks like today.