---
library_name: pytorch
tags:
  - pathology
  - vision-language
  - whole-slide-imaging
  - report-generation
  - ablation
---

# SlideFlame

This repository contains ablated variants of SlideFlame models for pathology
whole-slide image report generation.

SlideFlame is a slide-level vision-language model based on a Flamingo-style
architecture. It combines patch-level and slide-level visual features with a
language model to generate pathology reports from whole-slide image features.

The full project code is available on [GitHub](https://github.com/KatherLab/SlideFlame), while only the model weights 
are hosted here on Hugging Face.

## Model Weights

The checkpoint files are stored in the `weights/` directory:

| Variant | File |
| --- | --- |
| SlideFlame-GWI | `weights/slideflame-GWI.pt` |
| SlideFlame-GZI | `weights/slideflame-GZI.pt` |
| SlideFlame-LWI | `weights/slideflame-LWI.pt` |
| SlideFlame-LZI | `weights/slideflame-LZI.pt` |

## Variant Names

- `SlideFlame-GWI`: Global warm initialization variant
- `SlideFlame-GZI`: Global zero initialization variant
- `SlideFlame-LWI`: Local warm initialization variant
- `SlideFlame-LZI`: Local zero initialization variant

## Checksums

SHA-256 checksums are provided in `checksums.txt`.