# WorldShift AI — Create. Explore. Rescue.

## Inspiration
We wanted AI worlds people could explore, and an experience where intelligent navigation helps them work toward a concrete goal. WorldShift combines open-ended visual creation with hurricane rescue route-planning gameplay.

## What it does
Create Anything transforms a natural-language environment description into a high-quality first-person SDXL image. Waypoint then generates real action-conditioned video as the user moves with WASD and looks around. Users can save scenes and use lifecycle and recovery controls. A separate Rescue Mission challenges players to navigate hazards, rescue survivors and reach safety, with scoring and an A* route comparison. AI Showdown presents our existing PPO evaluation results.

## How we built it
Our local React application connects to a persistent Python GPU service on an RTX 4070 Ti SUPER. We use Hugging Face-hosted SDXL through Diffusers, Transformers, Accelerate and Safetensors for scene creation, and Overworld's Waypoint through official World Engine for interactive video. Native720 FP8 inference streams genuine JPEG frames over WebSockets with bounded queues, acknowledgments, session ownership and epoch handling. We publish a Hugging Face Space containing real outputs, model attribution and measured evidence.

## Challenges
Fast frame generation does not guarantee smooth browser presentation. We separately measured generated and canvas FPS, deliberate presentation skips, server drops and decode errors. Autoregressive visual drift can distort structures during longer movement. Clean scene restoration helps recover a view but does not prove persistent geography or guaranteed prompt fidelity. Shared-GPU enhancement and browser-dependent pointer lock also require care.

## Accomplishments
In a native720 fresh-seed recovery experiment, we measured 28.73 genuine generated FPS over 60 seconds, three successful recoveries and zero reported server drops, frame gaps or decode errors. Canvas presentation was 8.33 FPS in that historical run; later automated scheduling tests were more constrained. We preserved a deterministic rescue game and existing AI evaluations while building a real local generative-video pipeline.

## What we learned
Visual quality, responsiveness and honest measurements matter more than headline FPS. SDXL handles text; the tested Waypoint checkpoint is image- and action-conditioned. A* and PPO belong to the deterministic simulation, whose geometry is separate from the visual stream.

## What's next
Improve foreground browser presentation and experimentally validate prompt-grounded recovery. Keep the rescue experience reliable while treating generative exploration as an evolving creative prototype.

## Built with
React, TypeScript, Python, PyTorch, CUDA, Hugging Face Hub, Diffusers, Transformers, Accelerate, Safetensors, SDXL, Waypoint, FP8, WebSockets, A*, PPO.

## Links
Demo: https://huggingface.co/spaces/L3Harris/WorldShift-AI
Source: https://github.com/pranavsaigandikota/Morph

Submission draft only: review team names, event eligibility and final form fields before submission. No Devpost form has been submitted automatically.