EDITORIAL VIEW
Editor's Verdict
Oasis is a strong research record because the authors publish architecture, code, weights, and failure modes alongside the demo.
It is a technical demonstration, not a finished game. Its documented instability matters as much as the 20 FPS claim.
How It Works
- Oasis accepts keyboard input and generates visual state frame by frame without a conventional game engine. The authors describe a Transformer autoencoder and diffusion backbone.
What We Like
- The project publishes architecture, code, weights, and known failures.
- Keyboard conditioning makes the interaction model concrete.
- The page separates current results from future hardware projections.
Current Limitations
- The authors report distant fuzziness, temporal inconsistency, and domain-generalisation problems.
- Precise inventory and object control remain documented limitations.
- The 20 FPS figure is an official claim, not a site measurement.
Pricing & Access
- The project links a demo, source code, and model weights. Hosted-demo pricing and capacity are not publicly documented.
Verification Summary
- Video Runtime reviewed the official project page and linked code on 4 Sep 2026. No demo run was performed.
- Performance and architecture statements are labelled as official claims.
Technical notes & official performance
Official Performance Data
- Official claim: Oasis generates output at 20 frames per second.
- A 500M-parameter model is available for local use; hardware-dependent local results are not measured here.
Technical Notes
- The design combines a ViT-based spatial autoencoder with a DiT-style latent diffusion backbone.
- Frames are autoregressive and conditioned on user input; dynamic noising addresses error accumulation.
Sources
- Oasis project page ↗research · accessed 2026-09-04
- Oasis source code ↗github · accessed 2026-09-04
FROM PEOPLE WHO HAVE TRIED IT
User experiences
Reviewed community contributions, kept separate from our research and hands-on tests.
No published experiences yet. Tried this project? Share what worked and what didn’t.
Sharing experiences will open once email sign-in and submission protection are ready. You can still send a correction.
CONTINUE EXPLORING
Related projects & models
Selected by research links and category, not by popularity.
A world-model experience for creating and playing generated games through words, voice, and in-world actions.
Another frame streaming option in this section.Public product page; the play service is linked separatelyA source-available tool for running interactive generative-video pipelines, with a non-commercial alpha licence.
A related project selected in this entry’s research.Source-available, non-commercial; local or cloud GPU setup requiredExplore long-running autoregressive video generation and the infrastructure behind interactive streams.
Another frame streaming option in this section.Public repository; self-hosted setup and compatible GPU requiredA research baseline for streaming video generation that trains models on their own autoregressive outputs.
Another frame streaming option in this section.Public repository; self-hosted setup and compatible GPU requiredRead how official-source and hands-on records stay separate in our Methodology.