t2ct_cgv

Weights of Team CGV for the MICCAI 2026 VLM3D Challenge.

Code, setup, and inference instructions: github.com/nawooooojin/t2ct_cgv

File	Track	Submission ID
warp_ct_ema.ckpt	CT Volume Generation	26550529
warp_mr_ema.ckpt	MR Volume Generation	94215494
spacing_table.json, axcodes_table.json	MR Volume Generation (lookup tables)	94215494

The checkpoints hold the EMA weights of the denoising U-Net (unet.* and scale_factor) and are loaded by scripts/generate.py. Download everything with huggingface-cli download jasonna33/t2ct_cgv --local-dir weights.

• WARP-CT: latent diffusion with a 3D U-Net in the frozen Wan2.1 VAE latent space, conditioned on RadFinder text features.
• WARP-MR: the same model, with a (sequence, plane) class label and per-volume percentile intensity normalization.

EMA weights, rectified-flow sampling with CFG 5 and 100 steps.

Team

Team CGV: Woojin Na, Hyunjun Park, Jinah Park (KAIST).

License

CC BY-NC-SA 4.0. Pretrained models and datasets keep their own licenses.