---
title: DiffusionGemma Radiology VQA
emoji: 🩻
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 6.19.0
python_version: "3.12"
app_file: app.py
pinned: false
short_description: Radiology VQA & report infill with a discrete-diffusion LLM
startup_duration_timeout: 1h
---

# DiffusionGemma · Radiology VQA & Interactive Report Infill

Interactive demo for the paper
[*Discrete Diffusion Language Models for Interactive Radiology Report Drafting*](https://huggingface.co/papers/2607.01436).

It runs the LoRA finetunes from
[`gevaertlab/diffusiongemma-radiology-vqa`](https://huggingface.co/gevaertlab/diffusiongemma-radiology-vqa)
on top of the image-conditioned discrete-diffusion backbone
[`google/diffusiongemma-26B-A4B-it`](https://huggingface.co/google/diffusiongemma-26B-A4B-it)
(`DiffusionGemmaForBlockDiffusion`).

## What you can do

- **Visual Question Answering** — upload a medical scan (X-ray / CT / MRI), ask a
  question, get an answer. Pick the adapter trained on the dataset closest to your
  image style: VQA-RAD, SLAKE, or VQA-Med.
- **Bidirectional report infill** — the signature capability of discrete diffusion.
  Write a partial report with a `[BLANK]` marker; the model fills the hole using
  **both** the text before *and after* it (a true autoregressive model can only see
  the left context). This mirrors the paper's any-order infill sampler hook.

## How it works

DiffusionGemma generates text by iteratively denoising a fixed "canvas" of tokens with
bidirectional attention, rather than left-to-right one token at a time. The infill mode
clamps known token positions to fixed values every denoising step (per
[`models/infill.py`](https://github.com/mxvp/discrete_diffusion_RRG) in the reference
implementation) so the model only fills the masked span.

## Stack

- **Backbone**: `google/diffusiongemma-26B-A4B-it` (25.2B total / 3.8B active MoE).
- **Adapters**: `gevaertlab/diffusiongemma-radiology-vqa` (`diffusion-{vqarad,slake,vqamed}`).
- **Hardware**: ZeroGPU `xlarge` — the 26B checkpoint (~49 GB bf16) needs the full card.
- A custom `transformers` wheel providing the DiffusionGemma architecture is bundled in
  this repo and installed at runtime by `app.py` (Spaces installs `requirements.txt`
  before the repo files are copied in, so a local-path wheel can't be referenced there).

## Note

Research demo only — **not** a medical device and not for clinical use.