FlashWorld: High-quality 3D Scene Generation within Seconds

TL;DR: FlashWorld enables fast (7 seconds on a 1x A100/A800 GPU, 4 seconds on 1x H100/H800 GPU) and high-quality 3D scene generation across diverse scenes, from a single image or text prompt.

Demo

demo.mp4

🔥 News:

• [2026.2.6] Our paper is selected as an oral presentation in ICLR 2026!

• [2026.1.29] Our training code is released. Please fill in a Google form to get access to it.

• [2026.1.27] Our paper is accepted by ICLR 2026!

• [2025.10.19] We release a command line interface (CLI). If you want to achieve the best generation results with FlashWorld, you can create a good input JSON using the Web Interface and then use the CLI to regenerate and render the scene.

• [2025.10.19] We update the web interface for better logging and faster downloading.

• [2025.10.17] We release an online demo on Huggingface Spaces at FlashWorld Online Demo.

• [2025.10.16] Paper and local demo code released.

Installation

• install packages

pip install torch torchvision
pip install triton transformers omegaconf ninja numpy jaxtyping rich einops moviepy==1.0.3 accelerate opencv-python av plyfile ftfy pandas uvicorn nanobind

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

• install gsplat@1.5.2, diffusers@wan-5Bi2v and spz packages

pip install git+https://github.com/nerfstudio-project/gsplat.git@32f2a54d21c7ecb135320bb02b136b7407ae5712
pip install git+https://github.com/huggingface/diffusers.git@447e8322f76efea55d4769cd67c372edbf0715b8
pip install git+https://github.com/nianticlabs/spz.git@a4fc69e7948c7152e807e6501d73ddc9c149ce37

• (optional) install sage-attention package.

git clone https://github.com/thu-ml/SageAttention.git
cd SageAttention 
export EXT_PARALLEL=4 NVCC_APPEND_FLAGS="--threads 8" MAX_JOBS=32 # parallel compiling (Optional)
python setup.py install  # or pip install -e .

• clone this repo:

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

Local Web Interface

python app.py

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

If your machine does not have enough GPU memory, add the --offload_t5 and --offload_transformer_during_vae flags to offload text encoding to the CPU, which will reduce GPU memory usage with little impact on generation speed. You can also add the --offload_vae flag, which will greatly reduce GPU memory usage to below 10GB, but will significantly increase generation time. Please use this flag with caution.

On a single A800 GPU, the generation time and GPU memory usage under different settings are as follows:

Generation Time	GPU Memory	Flags
8.5s	51GB
16.6s	30GB	--offload_t5
20s	24GB	--offload_t5 --offload_transformer_during_vae
10min	9GB	--offload_t5 --offload_vae

Command Line Interface

python cli.py --input_dir /path/to/input/json/files --output_dir /path/to/output/directory --video --spz --ply

Parameters:

• --input_dir: Directory containing JSON files with generation parameters (required). We provide some examples in ./examples that you can use the directory directly.
• --output_dir: Directory to save generated results (required)
• --video: Generate video output
• --spz: Export results in SPZ format
• --ply: Export results in PLY format
• --video_fps: Video frame rate (default: 15)
• some flags are shared with the Local Web Interface

Note: The CLI interface provides better rendering results compared to the web interface, as it uses uncompressed Gaussian Parameters. Use the CLI for comparison if you want to use FlashWorld as your baseline.

More Generation Results

results.mp4

Clarification on WorldScore Evaluation

We identify two primary issues in the original WorldScore Evaluation framework:

1. Evaluation Frame Sampling Bias

• Issue: Metrics such as object_control, content_alignment, style_consistency, and subjective_quality were originally evaluated only on anchor views. This fails to provide a comprehensive assessment of the entire video sequence. Specifically, methods utilizing iterative inpainting often perform well on anchor views while producing incomplete or inconsistent results in intermediate frames.
• Solution: In our paper, we modified the sampling strategy from fixed anchor views to random frame sampling across segments. We have re-calculated the metrics for all 3D-based methods using this more rigorous standard.

2. Robustness of Camera Control Metrics

• Issue: The camera_control metric was found to be non-robust for two reasons:
  1. Inconsistency: Identical results can yield different numerical values across evaluation runs.
  2. Sensitivity: Despite our method achieving theoretically perfect camera control, the current metric fails to capture this specific advantage effectively.
• Solution: Consequently, we have omitted this specific metric from the final results presented in the paper.

To ensure a fair comparison with our work, please refer to the following evaluation metrics:

Table 1: Original WorldScore Scheme (Evaluated on Anchor Views)

We recommend using these metrics if you are developing video-based methods. Since re-evaluating all baselines can be computationally demanding, this set provides a standardized reference point using the original anchor view sampling.

Method	Camera	3D Consist.	Photo Consist.	Obj. Control	Content Align.	Style Consist.	Subj. Quality
WonderWorld	92.90	86.91	85.56	53.37	64.86	70.57	49.81
LucidDreamer	88.93	90.37	90.20	41.18	75.00	48.10	58.99
WonderJourney	84.60	80.60	79.03	37.10	35.54	62.82	66.56
Ours	84.43	85.87	86.72	50.28	56.54	79.36	52.75

---

Table 2: Revised Scheme (Evaluated on Random/Novel Views)

We recommend using these metrics if you are developing 3D-based methods. This scheme utilizes random frame sampling to more accurately reflect the consistency and quality of the entire generated 3D scene across all viewpoints.

Method	Obj. Control	Content Align.	Style Consist.	Subj. Quality
WonderWorld	52.09	56.82	75.92	41.28
LucidDreamer	43.48	59.41	66.41	48.02
WonderJourney	34.81	38.37	67.52	61.49
Ours	49.61	53.96	81.52	54.63

Statistical Summary (Mean Scores)

The table below summarizes the average scores across different evaluation versions:

Version	WonderWorld	LucidDreamer	WonderJourney	Ours
Original (Anchor View)	71.99	70.39	63.75	70.85
Revised (+Random View)	70.21	69.54	63.77	70.96
Paper (-Camera Control)	66.43	66.31	60.30	68.71

Citation

@inproceedings{
  li2026flashworld,
  title={FlashWorld: High-quality 3D Scene Generation within Seconds},
  author={Xinyang Li and Tengfei Wang and Zixiao Gu and Shengchuan Zhang and Chunchao Guo and Liujuan Cao},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=2IftRjRB07}
}

License

Licensed under the Apache-2.0 license.

If you have any questions, please contact me via imlixinyang@gmail.com.

Acknowledgements

This work is done with Hunyuan-World Team.