CylinderSplat: checkpoints and data

Release files of CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis (ICLR 2026, arXiv:2603.05882). Code: github.com/wangqww/CylinderSplat.

Reproduce the released results

From a clone of the code repository, in the environment of its README:

bash scripts/reproduce.sh --gpus 0          # or --gpus 0,1,2 to evaluate on several GPUs in parallel
HF_ENDPOINT=https://hf-mirror.com bash scripts/reproduce.sh --gpus 0      # through the hf-mirror.com mirror

The script downloads the checkpoints and the evaluation data below (about 20 GB), checks them against SHA256SUMS, extracts them and evaluates the five checkpoints.

Contents

path	size	contents
checkpoints/mp3d_stage1_pixel_256x512/model.safetensors	121 MB	pixel branch, Matterport3D, 256×512 (stage 1)
checkpoints/mp3d_stage2_volume_256x512/model.safetensors	121 MB	volume branch, Matterport3D, 256×512 (stage 2)
checkpoints/mp3d_stage3_joint_256x512/model.safetensors	121 MB	joint model, Matterport3D, 256×512 (stage 3)
checkpoints/mp3d_stage4_joint_512x1024/model.safetensors	121 MB	joint model, Matterport3D, 512×1024 (stage 4)
checkpoints/loc360_finetune_256x512/model.safetensors	121 MB	joint model fine-tuned on 360Loc, 256×512
checkpoints/pansplat_backbone/pansplat_last.ckpt	357 MB	the authors' PanSplat run at 256×512; stage 1 initialises its backbone from it, and the pixel model reads it when it is built
eval/pano_grf_evalsets_images.tar	0.3 GB	PanoGRF test and validation sets: rgb.png, rot.txt, tran.txt, depth.png per view
eval/pano_grf_evalsets_depth.tar	2.2 GB	the same views: depth_metric.npy, depth_conf.npy (UniK3D prior), depth_anywhere.png
eval/360Loc_atrium_images.tar	5.7 GB	360Loc atrium (the test scene): 360° images and camera_pose.json
eval/360Loc_atrium_depth.tar	11.4 GB	atrium: depth_metric/ (UniK3D prior), depthanywhere/
train/pano_grf_train_unik3d.part00.tar.xz … part09.tar.xz	146 GB in total	UniK3D prior of the Matterport3D training split, 2000 scenes per part
train/pano_grf_train_depth_anywhere.tar	3.4 GB	Depth Anywhere depth of the Matterport3D training split
train/360Loc_train_unik3d.tar.xz	17 GB	UniK3D prior of the 360Loc training scenes (concourse, hall, piatrium)
train/360Loc_train_depth_anywhere.tar	0.5 GB	Depth Anywhere depth of the same scenes (not read by training)
SHA256SUMS		SHA-256 of every file above

Every archive extracts inside its dataset root: the pano_grf_* archives into the PanoGRF folder (they hold png_render_* paths), the 360Loc_* archives into the 360Loc folder (they hold <location>/... paths). The .tar.xz files need tar -xJf (or xz -T0 -dc FILE | tar -x for parallel decompression).

The training images are not included: download pano_grf_lr.tar and the 360Loc release as described in PanSplat's Data Preparation, then extract the train/ archives into the same folders.

Sources and licenses

The checkpoints were trained by the authors of CylinderSplat and are released under the license of the code repository. Everything else is derived from third-party datasets and stays under their terms; it is provided only so that the published results can be reproduced. Using these files means accepting those terms.

• Matterport3D (project page): the renders come from the PanoGRF data preparation, as distributed by PanSplat. Matterport3D data may only be used under the Matterport3D Terms of Use (non-commercial academic use).
• Replica (Replica-Dataset) and Residential (SoftOcclusionMSI): test sets generated by PanoGRF; see their licenses.
• 360Loc (360Loc): images and poses of the atrium scene from the official release; see the 360Loc repository for its terms.
• Depth: depth_metric.npy / depth_metric/ are predictions of UniK3D (ViT-L), depth_anywhere.png / depthanywhere/ are predictions of Depth Anywhere, both made by the authors on the images above.

Rights holders who want a file removed can open an issue in the code repository.

Citation

@inproceedings{wang2026cylindersplat,
  title     = {CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis},
  author    = {Wang, Qiwei and Ze, Xianghui and Yu, Jingyi and Shi, Yujiao},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

Please also cite PanoGRF, PanSplat, Matterport3D, Replica, Residential (SoftOcclusionMSI), 360Loc, UniK3D and Depth Anywhere when you use the corresponding files.