sebasmos/latent-sr-embeddings

Latent-SR Embeddings: Precomputed VAE Latents for Medical Image Super-Resolution

Precomputed VAE latent embeddings from the paper:

"Domain-Specific Latent Representations Improve the Fidelity of Diffusion-Based Medical Image Super-Resolution"
Sebastian Cajas, Ashaba Judith, Rahul Gorijavolu, Sahil Kapadia, Hillary Clinton Kasimbazi, Leo Kinyera, Emmanuel Paul Kwesiga, Sri Sri Jaithra Varma Manthena, Luis Filipe Nakayama, Ninsiima Doreen, Leo Anthony Celi.
arXiv:2604.12152 (2026) — under review at Nature Scientific Reports.

📄 Paper: https://arxiv.org/abs/2604.12152
💻 Code: https://github.com/sebasmos/latent-sr

Dataset Description

Each directory contains precomputed encoder outputs (posterior mean, deterministic) for a given (VAE, dataset) pair. These are the latent inputs to the diffusion UNet — sharing them avoids re-encoding during training/evaluation.

Each .npy file is a single 2D latent of shape (C, H, W), stored as float32.

Directory Structure

<vae>_<dataset>/
├── train_latent/   # training split latents
├── valid_latent/   # validation split latents
└── test_latent/    # test split latents

VAE × Dataset Index

Directory	VAE	Latent shape	Dataset	Split sizes
medvae-4-3_brats	MedVAE (3×64×64)	(3,64,64)	BraTS 2023 brain MRI	train/val/test
medvae-4-3_cxr	MedVAE (3×64×64)	(3,64,64)	MIMIC-CXR chest X-ray	train/val/test
medvae-4-3_mrnet	MedVAE (3×64×64)	(3,64,64)	MRNet knee MRI	train/val/test
sdvae_brats	SD-VAE (4×32×32)	(4,32,32)	BraTS 2023	train/val/test
sdvae_cxr	SD-VAE (4×32×32)	(4,32,32)	MIMIC-CXR	train/val/test
sdvae_mrnet	SD-VAE (4×32×32)	(4,32,32)	MRNet	train/val/test
medvae-4-1_brats	MedVAE 4_1 (1×64×64)	(1,64,64)	BraTS	train/val/test
medvae-4-1_cxr	MedVAE 4_1 (1×64×64)	(1,64,64)	MIMIC-CXR	train/val/test
medvae-4-1_mrnet	MedVAE 4_1 (1×64×64)	(1,64,64)	MRNet	train/val/test
medvae-8-1_brats	MedVAE 8_1 (1×32×32)	(1,32,32)	BraTS	train/val/test
medvae-8-1_cxr	MedVAE 8_1 (1×32×32)	(1,32,32)	MIMIC-CXR	train/val/test
medvae-8-1_mrnet	MedVAE 8_1 (1×32×32)	(1,32,32)	MRNet	train/val/test
medvae-8-4_brats	MedVAE 8_4 (4×32×32)	(4,32,32)	BraTS	train/val/test
medvae-8-4_cxr	MedVAE 8_4 (4×32×32)	(4,32,32)	MIMIC-CXR	train/val/test
medvae-8-4_mrnet	MedVAE 8_4 (4×32×32)	(4,32,32)	MRNet	train/val/test
medvae-4-4_brats	MedVAE 4_4 (4×64×64)	(4,64,64)	BraTS	train/val/test
medvae-4-4_cxr	MedVAE 4_4 (4×64×64)	(4,64,64)	MIMIC-CXR	train/val/test
medvae-4-4_mrnet	MedVAE 4_4 (4×64×64)	(4,64,64)	MRNet	train/val/test

Usage

import numpy as np
from huggingface_hub import snapshot_download

# Download all embeddings for MedVAE on BraTS
local_dir = snapshot_download("sebasmos/latent-sr-embeddings")
latents = np.load(f"{local_dir}/medvae-4-3_brats/valid_latent/000001.npy")
print(latents.shape)  # (3, 64, 64)

Citation

@article{cajas2026domain,
  title   = {Domain-Specific Latent Representations Improve the Fidelity of Diffusion-Based Medical Image Super-Resolution},
  author  = {{Sebastian Cajas} and {Ashaba Judith} and {Rahul Gorijavolu} and {Sahil Kapadia} and {Hillary Clinton Kasimbazi} and {Leo Kinyera} and {Emmanuel Paul Kwesiga} and {Sri Sri Jaithra Varma Manthena} and {Luis Filipe Nakayama} and {Ninsiima Doreen} and {Leo Anthony Celi}},
  journal = {arXiv preprint arXiv:2604.12152},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.12152}
}

Code: https://github.com/sebasmos/latent-sr · Paper: https://arxiv.org/abs/2604.12152

License

MIT License.