license: apache-2.0
pipeline_tag: text-to-image
tags:
- medical
- retinal-imaging
- fundus
- ophthalmology
- text-to-image
- diffusion
library_name: pytorch
extra_gated_heading: Access to RetinaLogos Model Weights
extra_gated_description: >-
RetinaLogos is an open model released under Apache-2.0. To help us understand
usage and stay in touch with the community, please fill in the short form
below before downloading. Access is granted automatically.
extra_gated_prompt: >-
By requesting access, you confirm that the information provided is accurate
and agree to the terms below. Generated images are synthetic and must not be
used for clinical diagnosis.
extra_gated_fields:
Name: text
Affiliation: text
Country: country
Intended use:
type: select
options:
- Academic Research
- Educational Use
- Commercial Use
- Other
I will cite the RetinaLogos MICCAI 2025 paper in any work using this model: checkbox
I understand generated images are synthetic and will NOT use them for clinical diagnosis or patient care: checkbox
I accept the terms of the Apache License version 2: checkbox
extra_gated_button_content: Agree and Access Model
RetinaLogos: Fine-Grained Synthesis of High-Resolution Retinal Images Through Captions
RetinaLogos is a text-to-image framework that generates high-resolution Color Fundus Photographs (CFPs) from free-form textual descriptions of retinal conditions. Trained on RetinaLogos-1400k, a synthetic dataset of 1.4 million retinal caption–image pairs derived from Electronic Health Records (EHRs), it provides fine-grained semantic control over retinal anatomical structures, lesion types, and disease progression.
Published at MICCAI 2025 (LNCS 15975, pp. 474–484).
Model Highlights
- Fine-grained control: precise control over retinal anatomy and disease stages via free-form text prompts.
- Clinically plausible: 62.07% of text-driven synthetic images are indistinguishable from real CFPs by ophthalmologists.
- Downstream gains: 10%–25% improvement in diabetic retinopathy grading and glaucoma detection when used for augmentation.
- Architecture: built upon the Lumina-Next text-to-image framework (Apache 2.0).
Repository Contents
This repository hosts the stage-3 training checkpoint:
| File | Size | Description |
|---|---|---|
stage3_checkpoints7/consolidated.00-of-01.pth |
~6.1 GB | Consolidated model weights |
stage3_checkpoints7/model_args.pth |
1.5 KB | Model architecture arguments |
Usage
Inference and training code, environment setup, and detailed instructions are in the GitHub repository: uni-medical/retina-text2cfp (see codes/README.md). Download this checkpoint, point the inference script's model path to stage3_checkpoints7/, and follow the instructions there.
from huggingface_hub import snapshot_download
ckpt_dir = snapshot_download(repo_id="junzhin/RetinaLogo")
Intended Use & Limitations
This model is intended for research purposes — data augmentation, ophthalmology education, and the study of retinal disease representation. Generated images are synthetic and must not be used for clinical diagnosis or as a substitute for real patient imaging.
Citation
@InProceedings{NinJun_RetinaLogos_MICCAI2025,
author = { Ning, Junzhi and Tang, Cheng and Zhou, Kaijing and Song, Diping and Liu, Lihao and Hu, Ming and Li, Wei and Xu, Huihui and Su, Yanzhou and Li, Tianbin and Liu, Jiyao and Ye, Jin and Zhang, Sheng and Ji, Yuanfeng and He, Junjun},
title = { { RetinaLogos: Fine-Grained Synthesis of High-Resolution Retinal Images Through Captions } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {474 -- 484}
}
Acknowledgments
We thank the authors of Lumina-Next and Lumina 2.0 for their open-source frameworks.