I-Scene: 3D Instance Models are Implicit Generalizable Spatial Learners

🌟 CVPR 2026

🏠 Project Page | 📄 Paper | 🤗 Model | 📦 Dataset | 🎮 Demo

Lu Ling¹, Yunhao Ge², Yichen Sheng², Aniket Bera¹

¹Purdue University    ²NVIDIA Research

I-Scene reprograms a pre-trained 3D instance generator to act as a scene-level learner, replacing dataset-bounded supervision with model-centric spatial supervision. This unlocks the generator's transferable spatial knowledge, enabling generalization to unseen layouts and novel object compositions.

• Model Flexibility: A pre-trained 3D instance generator can be directly reprogrammed as a scene-level spatial learner, without scene-level annotations.
• Transferable Spatial Prior: The reprogrammed model's spatial prior provides a rich learning signal for inferring proximity, support, and symmetry from purely geometric cues.
• Data Independence: The model learns spatial knowledge on non-semantic scenes from randomly composed objects, removing dependency on annotated data.
• Strong Generalizability: It allows for easy generalization to unseen layouts and various spatial relations in a feed-forward manner without per-scene optimization.

• Release inference code and sparse structure flow transformer
• Release I-Scene-v1 inference checkpoint
• Release local Gradio demo
• Release HF online demo
• Release training data scripts
• Release evaluation code

Set up your virtual environment with CUDA 12.1.1 and Python 3.10. Then, install I-Scene with all inference and demo dependencies:

git clone https://github.com/LuLing06/I-Scene-project.git
cd I-Scene-project
pip install -e . -r requirements.txt

If you cannot have the virtual environment setup, you can still build most of the dependencies with the requirements.txt file. The most headache problem is installing nvdiffrast. You need nvcc installed and build nvdiffrast from scratch.

If you still have troubles, we have uploaded a Docker image we are using. See below.

• We provide a Dockerfile to build the Docker image.
• We also provide a pre-built Docker image on dockerhub: luling0506/iscene:latest.

Visit our Project Page for:

• Interactive 3D scene visualization
• Comparison with state-of-the-art methods
• More visualization examples

This repository includes a local Gradio demo for interactive segmentation and generation:

python interactive_demo.py --model LuLing/IScene

Note: The first run may be slow because the model checkpoint needs to be downloaded and cached.

This is different from a hosted web demo. A hosted online demo, such as a Hugging Face Space, will be released separately.

The public release loads the I-Scene checkpoint from Hugging Face:

LuLing/IScene

Minimal Python example:

from iscene.inference.inferencer import ISceneInferencer

inferencer = ISceneInferencer.from_pretrained("LuLing/IScene")
inferencer.infer_and_save_scene(
    scene_rgb_path="examples/Scenethesis/children_playroom2_rgb.png",
    instance_seg_path="examples/Scenethesis/children_playroom2_seg.png",
    output_dir="outputs/example_gs",
    only_3dgs=True,
    seed=43,
)

Run Gaussian Splatting output on a bundled example:

python run_inference.py \
  --model LuLing/IScene \
  --rgb examples/Scenethesis/children_playroom2_rgb.png \
  --mask examples/Scenethesis/children_playroom2_seg.png \
  --output_dir outputs/example_gs \
  --only_3dgs

Run GLB mesh export:

python run_inference.py \
  --model LuLing/IScene \
  --rgb examples/Scenethesis/children_playroom2_rgb.png \
  --mask examples/Scenethesis/children_playroom2_seg.png \
  --output_dir outputs/example_glb

Important options:

--model       I-Scene model id or local model package path. Default: LuLing/IScene
--base_model  Optional TRELLIS base model id or local mirror path.
--seed        Random seed. Default: 43
--simplify    Mesh simplification ratio for GLB export. Default: 0.95
--save_dbg    Save debug artifacts.
--only_3dgs   Skip GLB export and save Gaussian PLY files only.

Typical outputs:

scene_pred.ply                 # merged Gaussian Splatting scene
scene_pred.glb                 # merged mesh scene, if GLB export is enabled
instance_XX.*                  # per-instance asset for mask label XX

If you find this work helpful, please consider citing our paper:

@article{ling2025iscene,
  title={I-Scene: 3D Instance Models are Implicit Generalizable Spatial Learners},
  author={Ling, Lu and Ge, Yunhao and Sheng, Yichen and Bera, Aniket},
  journal={arXiv preprint arXiv:2512.13683},
  year={2025}
}

This project is licensed under the MIT License - see the LICENSE file for details.

I-Scene builds on the TRELLIS image-to-3D backbone and uses microsoft/TRELLIS-image-large as the base model. Please preserve upstream TRELLIS attribution and license terms when redistributing derived code or model packages.

We thank the authors of TRELLIS, and other related works for their inspiring research.

The website template is borrowed from Nerfies and MIDI.