README.md

metadata

title: FlashWorld
emoji: 🌎
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.49.1
app_file: app_gradio.py
pinned: false
license: cc-by-nc-sa-4.0
python_version: 3.10.13

FlashWorld: High-quality 3D Scene Generation within Seconds

TL;DR: FlashWorld enables fast (7 seconds on a single A100/A800 GPU) and high-quality 3D scene generation across diverse scenes, from a single image or text prompt.

FlashWorld also supports generation with only 24GB GPU memory!

Demo

https://github.com/user-attachments/assets/12ba4776-e7b7-4152-b885-dd6161aa9b4b

🔥 News:

We are actively building an online demo. Please stay tuned!

• [2025.10.16] Paper and local demo code released.

Installation

• install packages

pip install torch torchvision
pip install triton transformers pytorch_lightning omegaconf ninja numpy jaxtyping rich tensorboard einops moviepy==1.0.3 webdataset accelerate opencv-python lpips av plyfile ftfy peft tensorboard pandas flask

Please refer to the requirements.txt file for the exact package versions.

• install [email protected] and diffusers@wan-5Bi2v packages

pip install git+https://github.com/nerfstudio-project/gsplat.git@32f2a54d21c7ecb135320bb02b136b7407ae5712
pip install git+https://github.com/huggingface/diffusers.git@447e8322f76efea55d4769cd67c372edbf0715b8

• clone this repo:

git clone https://github.com/imlixinyang/FlashWorld.git
cd FlashWorld

• run our demo app:

Local Demo (Flask + Custom UI):

python app.py

ZeroGPU Demo (Gradio):

python app_gradio.py

If your machine has limited GPU memory, consider adding the --offload_t5 flag to offload text encoding to the CPU, which will reduce GPU memory usage. Note that this may slow down the generation speed somewhat.

Then, open your web browser and navigate to http://HOST_IP:7860 to start exploring FlashWorld!

ZeroGPU Deployment

This repository is compatible with Hugging Face Spaces using ZeroGPU. The app_gradio.py file provides a Gradio interface with:

• 15-second GPU budget per generation (configurable via @spaces.GPU(duration=15))
• Model loading happens outside the GPU decorator for efficiency
• Supports both image and text prompts
• Camera trajectory input via JSON
• Outputs 3D Gaussian Splatting PLY files

To deploy on Hugging Face Spaces:

1. Create a new Space with ZeroGPU hardware
2. Set app_file: app_gradio.py in the README header
3. The model checkpoint will be automatically downloaded from HuggingFace Hub

More Generation Results

https://github.com/user-attachments/assets/bbdbe5de-5e15-4471-b380-4d8191688d82

License

Licensed under the CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International)

The code is released for academic research use only.

If you have any questions, please contact me via [email protected].