Scaling Video Pretraining with Imagination Models

Photon-1, a 106B-A5B MoE transformer trained on 18 years of computer demonstration video using imagination models, outperforms a production LLM trained with 30x more FLOPs on internal computer use benchmarks and is 3x cheaper to serve. Developed by David Li and Jonathan Li, the model learns from observation alone by predicting future latent states, and despite seeing only computer use video, it can learn to play checkers and simulate billiard physics better than LLMs.

Scaling Video Pretraining with Imagination Models David Li & Jonathan Li · Jul 23, 2026 Internet video contains millions of hours of people using computers, acting in the physical world, interacting with one another, and performing skilled work. We introduce imagination models , a simple foundation model architecture that unlocks scalable learning from this internet video. These models learn to imagine the future in a learned representation space. During pretraining, imagination models implicitly learn to act despite seeing no action labels. Afterwards, reinforcement learning can be scaled to consistently improve their competence. We test the imagination model architecture with Photon-1, a sparse 106B-A5B MoE transformer trained on 18 years of computer demonstration video. On internal computer use benchmarks, Photon-1 outperforms a production LLM trained with 30x more FLOPs, and is 3x cheaper to serve . And despite seeing only computer use video, it serves as a basis for broader world understanding; for example, it can learn to play checkers and simulate billiard physics better than LLMs. It also demonstrates desirable human behavior present in pretraining data, such as steering AI tools to complete tasks. Learning from observation Humans can learn by watching other people. What we learn from is often not the raw actions a person takes, but higher level concepts of what the person is doing . Video contains an enormous amount of this observational data, which imagination models can learn from. Our key result is that predicting future states can teach models to complete tasks despite never seeing actions during pretraining, allowing them to learn broadly from video through observation alone. Photon-1 is our initial test of imagination models, trained on only computer use videos. We use computer environments as a testbed because they are open-ended, cheap to run at scale, and provide access to a large share of humanity's knowledge and work. Training imagination models Imagination models predict future frames autoregressively using a next- latent -token-prediction objective. They do not generate pixels during pretraining; everything is modeled in representation space. Our approach is made possible by our vision encoder, which uses finite scalar quantization FSQ to compress each frame into 960 discrete tokens. Each token is an 8‑dimensional vector, where each dimension takes one of five values {−1, −1/2, 0, 1/2, 1}, yielding a vocabulary of 5⁸ possible codes. The encoding totals just 2.2KB, providing 100× better compression compared to existing OCR and multimodal-model representations , while preserving the text, layout, and state changes needed to understand video. To achieve this rate of compression, Photon-1 uses a differential latent encoder, which encodes video frames as pairs to allow the latents to describe differences between frames. Our pretraining corpus originates from an internal index of 2 billion publicly available videos, filtered down to approximately 2 million computer screen recordings. We performed extensive video and frame-level filtering to remove non-computer-use content, and built an internal keyframe detection model to remove redundant frames. We pretrained Photon-1 from scratch for a single epoch on the resulting 575-million-frame dataset, equivalent to 552 billion tokens or 18 years of video sampled at 1 frame per second. Pretraining the 106B-A5B MoE transformer at 32K context length required approximately 30,000 H200 GPU-hours, corresponding to 4.4×10²² training FLOPs. We implemented training in PyTorch with custom fused kernels for our vision encoder and MoE layers, sustaining 40% end-to-end MFU. After training Photon-1, we finetuned an open-source single-stream diffusion transformer as an auxiliary image generation model to convert Photon-1’s representation space to a screenshot image for visualization purposes. This decoder is used only for visualization and is not part of Photon-1’s training or inference. Acting from imagination We show that imagination models learn implicit policies. We finetuned Photon-1 on small amounts of computer use data <35k trajectories to teach it the correct action and instruction format. Under our computer use setup, Photon-1 imagines what should happen first, then generates the action to get to that state. We applied online reinforcement learning to the resulting finetuned model. Our online reinforcement learning engine performs real-time rollouts on virtual machines at scale, and programmatically verifies the outcomes on the machines to provide a reward signal. On Linux virtual machines, we augment OS functionality with different desktop environments LXQt, Xfce, MATE, GNOME, and Plasma . Each environment has a unique Google account sign-in for login-restricted web applications and contains an internal clone of ChatGPT without rate limits. Our preliminary results below show that scaling reinforcement learning brings Photon-1's computer use capability beyond a production LLM Gemini 3.1 Flash-Lite on internal benchmarks, while using less compute. We compare against LLMs to show that imagination models can be competitive, despite being trained on at least 30× less compute and being 3× cheaper to run, compared to Gemini 3.1 Flash-Lite: | Model | Pretraining compute | Weighted inference cost / 1M tokens | |---|---|---| | Gemini 3.1 Flash-Lite | | $0.36 | Photon-1 | $0.11 | Interestingly, we also observe that the predicted representations after RL become difficult to visualize and understand. In combination with improvements in task execution, this suggests that Photon-1 is honing its use of generated latent tokens to improve planning and reasoning. Simulating the desktop After a small amount of action-conditioned finetuning, Photon-1 can simulate the consequences of actions in desktop environments. In the examples below, we provide an initial state and an action, then Photon-1 generates the latent representation of the resulting state. Summary: Create a Python file in VS Code and write a hello world script. Generalizing beyond the desktop We show that Photon-1’s representations transfer beyond computer use, despite only being pretrained on computer demonstration video. To test whether imagination pretraining provides useful priors for learning new dynamics, we finetuned the model on games and physical simulations not present during pretraining. We compared Photon-1 against a strongly initialized baseline using the same vision encoder, and a pretrained LLM with similar size to Photon-1. We finetuned all three models on 20,000 tournament checkers games and evaluated them on two axes: world simulation predicting how the board evolves and skill the quality of the moves they choose . Photon-1 outperforms both baselines on both axes. We then repeated the comparison on physical dynamics, finetuning on 10,000 synthetically generated billiard games to simulate positions of balls at 5 frames per second. Photon-1 again outperformed both baselines in simulation accuracy. Imitating human behavior Photon-1 learns human priors present in our pretraining data, such as using AI tools to complete tedious tasks. For example, after reinforcement learning, our model learns to use our internal ChatGPT clone to generate documents and answer knowledge-related questions. Task: Two rocky planets have the same composition as Earth. One is about five times Earth's mass, while the other is half Earth's mass. Which planet should have the higher average density and why? Pretraining on human interactions with AI tools teaches our models how to use, guide, and correct LLMs as tools. Related work We think that imagination models solve some core problems with existing approaches that use video to learn a policy. The standard method to learn from video demonstrations is to infer the actions taken in a video using an inverse dynamics model IDM , then train a model to predict the next inferred action. However, this makes internet-scale pretraining dependent on the capabilities of the IDM, since a video is only useful when the IDM can describe its actions accurately. Improving that coverage requires specialized labeled data for each application, weakening the core advantage of internet-scale pretraining. More fundamentally, IDMs require actions you can infer and write down. Much of what video can teach has no simple action representation at all. Compared to existing world models and video diffusion models, we show that imagination models can complete complex tasks that require planning and an understanding of text and layout, such as computer use. This is a hard bar that diffusion models struggle to cross. Our method is able to complete these tasks while being simple and easy to scale. Towards observational learning Photon-1 is our initial test of imagination models. We expect both pretraining and reinforcement learning to continue improving with scale. In principle, imagination models are not limited to learning from just computer use video or demonstrations with easily identifiable actions. We plan to extend pretraining to broader observational data, such as physical dynamics, skilled labor, and social interaction. Scaling imagination models opens a path to intelligence that learns generally by observing the world directly, without a human first translating it into text. We see this as the foundation for agents that can learn autonomously by interacting with the world. We’re a team of two based in San Francisco. We’re always looking for exceptional thinkers, engineers, and researchers. If our work sounds interesting to you, leave us a note at team@inductionlabs.com mailto:team@inductionlabs.com .