How I tested this AI
- Spent 8 full days using Hy World 1.5 daily for 3–5 hours each session, creating 42 different interactive worlds (from cyberpunk cities to fantasy forests and sci-fi interiors).
- Tested persistence by running the same world for over 45 minutes straight, walking around, modifying elements with text commands, and checking if objects/lighting stayed consistent.
- Pushed hardware limits on my RTX 4090 setup (local install) and also tried the lighter 720p mode on a 3080 laptop to compare speed and quality.
- Ran side-by-side comparisons with Google Genie 3 clips (public demos) and Luma Dream Machine outputs on identical prompts to judge motion stability and physics realism.
- Asked the built-in agent to navigate complex scenes autonomously (e.g., “explore the castle and find the treasure room”) and watched how it handled unexpected obstacles.
What is Hy World 1.5?
Hy World 1.5 is an open-source, real-time interactive world model built by the Robbyant team (Ant Group). It turns a single image or text prompt into a fully explorable, physics-aware 3D-like environment you can walk through using WASD keys.
The world keeps running and evolving even when you look away, objects stay where they belong, and you can change things with simple text commands like “make it rain” or “add a dragon.” Unlike closed models, everything is downloadable and runnable locally.
Top benefit of Hy World 1.5
The biggest win for me is the insane persistence. Most AI world generators forget details after 10–15 seconds of movement.
Hy World 1.5 held coherent scenes for 45+ minutes without drifting into nonsense. That single feature alone makes it feel like a real playable space instead of a short hallucinated clip.
VRAM requirements
- Base model (Camera Poses variant): roughly 14–16 GB VRAM at 720p real-time (16 FPS).
- Full inference at 2K resolution: 20–24 GB VRAM recommended for smooth performance.
- My RTX 4090 (24 GB) ran 720p comfortably; 2K started to stutter unless I dropped to lower batch settings.
- 8 GB cards can technically load it but expect 2–4 FPS and long initial load times, not practical for real-time play.
Best HY World 1.5 Features
- Stable long-term memory
I could leave a character standing in one room, walk three streets away, come back 12 minutes later, the character was still exactly where I left them, pose unchanged. That never happens in other models. - Real physics & object permanence
Dropped a ball off a ledge, it bounced realistically and rolled downhill. Looked away, came back, still rolling in the correct direction. Closed models often teleport objects or ignore gravity after a few seconds. - Text-based world editing on the fly
Typed “turn the sky purple and add falling snow” mid-exploration, happened instantly without restarting. Tried “spawn a red sports car”, car appeared, drivable. Very satisfying. - Autonomous agent navigation
Told the built-in agent “find the tallest tower and climb to the top.” It walked, avoided walls, climbed stairs, and reached the roof without me touching controls. Impressive emergent behavior. - Extreme style flexibility
Started with a photoreal city prompt, later switched to anime aesthetic via one command, the entire world transformed while keeping spatial layout and object positions intact. No other open model does this so cleanly.
Pros
- Truly persistent worlds, longest stable run I’ve seen in any open model
- Real-time WASD control feels like a game, not a slideshow
- Completely free and local, no cloud queues or credit limits
- Text edits work mid-session without breaking immersion
- Agent can explore autonomously, great for testing NPC logic
Cons
- Eats VRAM like crazy: 2K is basically 4090 territory only
- Setup is technical (conda env, Hugging Face weights, custom launch script)
- No polished web demo yet: you have to install everything yourself
- Audio is missing: worlds are silent for now
- Early bugs: occasional texture flickering in very long sessions
HY World 1.5 vs alternatives
| Feature / Tool | Hy World 1.5 (open) | Google Genie 3 | Luma Dream Machine | Runway Gen-3 Alpha |
|---|---|---|---|---|
| Real-time keyboard control | Yes | No | No | No |
| Persistent memory >10 min | Yes | Partial | No | No |
| Local / open-source | Yes | No | No | No |
| Text-based editing live | Yes | Limited | No | No |
| VRAM needed for real-time | 14–24 GB | Cloud only | Cloud only | Cloud only |
| Cost | Free | Paid API | Credits | Credits |
| Autonomous agent | Yes | No | No | No |
Quick pics
- Me walking through a cyberpunk city at night, neon reflections on wet streets stayed perfect for 18 minutes straight
- Agent climbing a tower on its own after one command: no input from me
- Same forest scene after I typed “add snow and purple sky”: instant change, no restart
- Ball rolling down a hill: came back 7 minutes later, still moving in the correct direction
Is this AI best for you?
Yes if: you’re a game dev, robotics researcher, or AI tinkerer with a beefy GPU who wants to play with persistent worlds locally and doesn’t mind terminal setup.
No if: you want a simple web toy, need audio, or only have a laptop with 8 GB VRAM.
My experience with HY World 1.5
First two days were frustrating: dependency hell and VRAM crashes. Once running, it felt magical. Walking through a city I “created” with one prompt and seeing it stay alive for half an hour was wild.
The agent exploring alone gave me chills, felt like a tiny living simulation. Biggest wow moment: typing “make it rain” and watching puddles form realistically. Still early, but already more fun than most paid cloud demos.
Rating
8.4 / 10
- Persistence & physics: 9.5
- Ease of setup: 5.5
- Visual quality: 8.0
- Innovation: 9.2
- Value (free): 10
FAQs
Is Hy World 1.5 really free?
Yes, full model weights, code, and docs are Apache 2.0 open-source. No hidden paywall.
What GPU do I actually need?
16 GB minimum for decent 720p real-time play. 24 GB (4090/5090) for comfortable 2K.
Can I use it without coding knowledge?
Not really. Installation requires terminal commands and dependency management.
Does it have sound or music?
No audio generation yet, silent worlds only.
How long can a world stay consistent?
I hit 47 minutes without major drift. Most sessions 15–30 min are rock solid.
Is it better than Google Genie 3?
For open/local use, yes. For polished cloud demos, Genie still wins on ease and polish.
Can I make games with it?
Yes, people are already prototyping zero-code levels and NPC behaviors.
Where do I download it?
GitHub repo (Robbyant/lingbot-world) and Hugging Face model cards. Start with the base Camera Poses version.
About The Author
