mickmumpitz/Wan2.1-Pano360-LoRA

Pano 360 Video LoRA (ComfyUI / WAN 2.1) — for ComfyUI-SplatKit

ComfyUI-ready conversion of Matrix-3D's pano_video_gen_720p LoRA. This is the panoramic (360° / equirectangular) video-generation LoRA that drives the flythrough in ComfyUI-SplatKit: it teaches WAN 2.1 I2V to generate a coherent camera move through a 360° scene and to fill the black disocclusion holes the mesh-render step leaves behind.

• File: pano_video_gen_720p_comfy.safetensors (BF16, ~292 MB)
• Base model: WAN 2.1 I2V 14B (720P)
• Model folder: ComfyUI/models/loras/

What this is (and isn't)

This is not a retrain. It is a lossless key-format conversion of Skywork's original weights from DiffSynth/PEFT naming (blocks.*.lora_A/lora_B.default.weight) into the ComfyUI/WAN convention (diffusion_model.blocks.*.lora_down/lora_up.weight) that ComfyUI loads natively — the same convention the lightx2v distill LoRAs use. No tensor values are altered; only key names change and weights are stored as BF16 to match the WAN runtime.

The conversion was done with tools/convert_pano_lora.py in the SplatKit repo and structurally validated against a known-good ComfyUI WAN LoRA (every pano-targeted module is present).

License & attribution

This conversion is redistributed under the MIT License, unchanged from the source.

Original weights © 2025 Skywork AI — Matrix-3D (pano_video_gen_720p), MIT License. This repository redistributes those weights in a re-keyed form; all credit for the model itself belongs to the Matrix-3D authors.

The LoRA is a fine-tune of Wan2.1 (Apache-2.0). Use of the produced videos is subject to the terms of both upstream projects — both permit commercial use with attribution.

📚 Citation:

@article{yang2025matrix3d,
  title   = {Matrix-3D: Omnidirectional Explorable 3D World Generation},
  author  = {Yang, Zhongqi and Ge, Wenhang and Li, Yuqi and Chen, Jiaqi and
             Li, Haoyuan and An, Mengyin and Kang, Fei and Xue, Hua and Xu, Baixin and
             Yin, Yuyang and Li, Eric and Liu, Yang and Wang, Yikai and
             Guo, Hao-Xiang and Zhou, Yahui},
  journal = {arXiv preprint arXiv:2508.08086},
  year    = {2025},
  url     = {https://arxiv.org/abs/2508.08086}
}

Acknowledgements

• Skywork AI / Matrix-3D — the original pano_video_gen_720p LoRA weights.
• Wan-Video / Wan2.1 — the base image-to-video model.
• Packaged for ComfyUI by mickmumpitz as part of ComfyUI-SplatKit.