WorldFoundry
English中文

FlashWorld

Method overview: a dual-mode multi-view latent diffusion model distilled across generation modes.

FlashWorld from Xiamen University / Tencent: fast 3D scene generation from image or text; Verified.

Edit this pageGitHubMarkdown

FlashWorld generates 3D scenes in seconds: from one image (plus optional prompt) it produces a camera-navigable 3D scene rather than a flat clip. In this catalog the in-tree pipeline, operator, and representation runtime are exposed and runner status is Verified for the image-conditioned camera-scene demo with local FlashWorld/Wan assets.

Official paper teaser: FlashWorld generates high-quality 3D scenes in seconds across diverse environments.
Official paper teaser: FlashWorld generates high-quality 3D scenes in seconds across diverse environments.
Method overview: a dual-mode multi-view latent diffusion model distilled across generation modes.
Method overview: a dual-mode multi-view latent diffusion model distilled across generation modes.

The recorded paper is FlashWorld: High-quality 3D Scene Generation within Seconds (2025).

Project page: imlixinyang.github.io/FlashWorld-Project-Page. Code: imlixinyang/FlashWorld.

Paper: FlashWorld: High-quality 3D Scene Generation within Seconds (arXiv 2510.13678). Project: imlixinyang.github.io/FlashWorld-Project-Page. Code: imlixinyang/FlashWorld (Apache-2.0). Weights: imlixinyang/FlashWorld — Hugging Face model metadata is CC-BY-NC-SA-4.0, so honor the artifact-specific license. Institutions: Xiamen University / Tencent.

The stack distills a Wan-based video generator into a fast Gaussian-style scene generator. Runtime needs FlashWorld, Wan2.2 (Wan-AI/Wan2.2-TI2V-5B-Diffusers), diffusers / transformers, and CUDA assets.

Build your run

Pipeline
FlashWorldPipelineworldfoundry.pipelines.flash_world.pipeline_flash_world
TaskOfficial Camera Scene official-camera-scene
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  flashworld \
  --pipeline.task-profile official-camera-scene \
  --json

What it is for

flashworld is the recorded reconstruction line. Start here to emit geometry from the documented views.

FlashWorld distills a Wan-based video generator into a fast 3D-scene generator with Gaussian-style scene representations (per arXiv 2510.13678). Execution requires FlashWorld, Wan2.2, diffusers/transformers, and CUDA assets; no empty runner shim is registered. It is published by Xiamen University / Tencent.

Catalog tasks include fast 3D scene generation from image or text.

Recorded evaluation entry points include WorldScore (Standard scoring for navigable generated scenes).

Runtime contract

Optional input_path, prompt, interactions, num_frames, and fps; artifacts generated_3d_asset, generated_video, and flashworld.ply.

CLI task profiles: official-camera-scene. WorldFoundry binds FlashWorldPipeline.

Family status: Verified.

Recorded official-camera-scene defaults: num_frames=16, fps=15.

Execution remains checkpoint/GPU/dependency gated — stage all recorded assets first.

Running in WorldFoundry

Launch worldfoundry-eval run flashworld --pipeline.task-profile official-camera-scene to write geometry. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

Recorded weight repositories include imlixinyang/FlashWorld and Wan-AI/Wan2.2-TI2V-5B-Diffusers. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh flashworld. Add --download to fetch missing public checkpoints.

Status: Verified.