---
license: apache-2.0
library_name: diffusers
pipeline_tag: unconditional-image-generation
tags:
  - zoomldm
  - cdm
  - dit
  - histopathology
  - brca
  - custom-pipeline
widget:
  - src: demo_images/input.jpeg
    prompt: Sample BRCA conditioning embedding (magnification class 0)
    output:
      url: demo_images/output.jpeg
---

> [!WARNING] we do not have a full checkpoint conversion validation, if you encounter pipeline loading failure and unsidered output, please contact me via bili_sakura@zju.edu.cn

# BiliSakura/ZoomLDM-CDM-brca

Diffusers-style wrapped **CDM (DiT)** checkpoint for BRCA, converted from ZoomLDM `cdm_dit` training outputs.

## Model Description

- **Architecture:** DiT-B style conditioning diffusion model (CDM)
- **Domain:** BRCA conditioning space used by ZoomLDM
- **Output:** conditioning tokens/embeddings (`(B, 512, 65)`)
- **Format:** custom diffusers pipeline (`pipeline.py`)

## Intended Use

Use this model to sample BRCA conditioning embeddings that can be consumed by downstream ZoomLDM workflows.

## Out-of-Scope Use

- Not a complete pixel-space generator by itself.
- Not intended for clinical or diagnostic use.
- Not validated for non-BRCA domains without adaptation.

## Files

- `pipeline.py`: custom `DiffusionPipeline` implementation (`CDMDiTPipeline`)
- `model_index.json`: diffusers metadata
- `cdm/`: active model weights/config used by pipeline
- `scheduler/`: DDIM scheduler config
- `model_raw.safetensors`: non-EMA training weights (optional)
- `optimizer.pt`: optimizer state (optional)
- `config.json`: conversion metadata

## Usage

```python
import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/ZoomLDM-CDM-brca",
    custom_pipeline="pipeline.py",
    trust_remote_code=True,
).to("cuda")

out = pipe(
    batch_size=2,
    magnification=torch.tensor([0, 0], device="cuda"),  # class labels 0..7
    num_inference_steps=50,
    guidance_scale=1.0,
)

samples = out.samples  # (B, 512, 65)
```

## Limitations

- Produces conditioning embeddings, not final images.
- Requires correct class/magnification label conventions.
- Inherits data biases and quality limits from the original training data.

## Citation

```bibtex
@InProceedings{Yellapragada_2025_CVPR,
  author = {Yellapragada, Srikar and Graikos, Alexandros and Triaridis, Kostas and Prasanna, Prateek and Gupta, Rajarsi and Saltz, Joel and Samaras, Dimitris},
  title = {ZoomLDM: Latent Diffusion Model for Multi-scale Image Generation},
  booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
  month = {June},
  year = {2025},
  pages = {23453-23463}
}

@inproceedings{Peebles2023DiT,
  title={Scalable Diffusion Models with Transformers},
  author={Peebles, William and Xie, Saining},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2023}
}
```