GTA world model

GTA world model

Makes video from one start image while you hold keys, frame after frame, streamed to a browser. Weights and inference code for a world model trained on GTA5 material, for researchers and ML engineers.

Made by lucidml on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is an interactive world model that makes video from a single starting image. You hold keys as you play, and the model produces the footage that follows, frame after frame, streamed to a browser. The author, lucidml, designed the architecture, called Summaryformer, and trained it from scratch. This checkpoint learned from GG800, an 800-hour collection of GTA5 material.

Who it suits: It suits researchers who study video generation, world models or long memory in transformers, and engineers who want weights together with a working streaming runtime. Curious players can try the online demo the card links.

Industry | Who uses it | How they use it | Example
Research | Video and world model researchers | Study the Summaryformer memory design | Read how older frames are kept only as short summaries
Software | Machine learning engineers | Run the weights with the included runtime | Look at the sampler and the streaming server
Education | Students of transformers | Learn how attention can read summaries | Compare 6528 keys in memory against 26208 without summaries
Entertainment | Game and interactive media teams | Watch a generated scene respond to keys | See how a model handles continuous play

If you need training code, skip it, because the repository holds none.

Where to use it: The model is used in these places.

• A web browser: the card links an online demo where you give a start image and hold keys.
• The Hugging Face model page: read the card and fetch the files from there.
• Python on Windows, macOS or Linux: download the weights and runtime to your own computer.

Use case | Who | What they get back
Play a generated scene | Anyone curious | Video that follows the keys you hold
Study the memory design | Researchers | A written account of how frames are stored and read
Run the weights yourself | Engineers | The weights plus the model, sampler and streaming server code

How to set it up: You need a few things first.

• Python and pip
• A terminal on Windows, macOS or Linux
• Hardware: the card text available here names no GPU or memory need

These are Hugging Face's steps for getting the files.

1. Install the Hugging Face hub package with pip.
2. Download the model's files with the hf command.
3. Follow the run steps in the model card on its Hugging Face page.

pip install -U huggingface_hub
hf download lucidml/gta

How to use it: In the online demo, give the model one starting image and hold keys while it makes the next frames. A local copy holds the model, its memory, the sampler and a streaming server, which sends frames to a browser as you play. The card text available here is cut off before the run steps, so check the model page for the exact commands.

Example: You start with one image of a scene and begin holding keys. The model builds the footage that follows, and your browser plays it as it arrives. Older frames reach the model only as short summaries, which keeps its memory small.

For newcomers and experts: An interactive world model makes video that responds to what you do. Here you give one starting picture, hold keys, and watch the video appear as it is made. The first small thing to try is the online demo the card links.

Each frame is stored as a record. It holds 1008 raw tokens at 384x672, followed by 96 summary tokens that the same 28 blocks write in the same pass, with no separate summarizer. A frame attends to itself, the last 2 frames in full, older frames inside the window as summaries only, and the start image in full. The trade-off is that old frames survive only as outlines. The trained window is 24 frames and the server default is 16.

You give the model one start image
You hold keys while you play
The model generates each next frame
The server streams the frames to a browser

Good to know: The memory numbers below come from the card.

Title: Keys held in the memory buffer
With summaries: 6528 keys
Every frame at full size: 26208 keys

• Contents: the repository holds the weights and the inference runtime, with no training code.
• Training data: this checkpoint learned from one dataset, GG800.
• License: apache-2.0, as the record states.
• Cost to run: none named in the text available here. The source lists no API key, account or paid service.
• Data it reaches: it downloads files from Hugging Face, and the streaming server sends frames to a browser. The source does not say what the online demo collects.
• Not for: training a new model, since no training code is included.

On Hugging Face, 13 November 2026

• Task: image to video
• Downloads in the last 30 days: 0
• Likes: 11
• License: apache-2.0
• Tags: world-model, interactive, real-time, autoregressive, video-generation, image-to-video

Links

• Model: https://huggingface.co/lucidml/gta
• Files: https://huggingface.co/lucidml/gta/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lucidml/gta

Then follow the usage in its model card. Nothing in this kit runs by itself.