ZoomLDM

A pathologist reads a slide by changing objectives. At 20x a field shows nuclear atypia and mitotic figures; at 1.25x the same tissue resolves into a tumor's growth pattern and the margin where it meets stroma. A whole-slide image can run to 32,000 x 32,000 pixels, far beyond what a diffusion model can be trained on directly, so generative models for histopathology have instead learned to draw small fixed-size patches cropped at one magnification. Such a model has no representation at all of the architecture that only becomes visible three zoom levels out.

ZoomLDM, from Stony Brook University's computer vision lab and presented at CVPR 2025, treats magnification as an explicit conditioning variable rather than a fixed property of the training crop. One set of weights generates 256 x 256 patches at eight magnifications spanning 20x down to 0.15625x, where a single patch covers an entire slide. Sharing weights across scales is what makes the coarse end workable: the 0.15625x level has only about 2,500 training patches, and a standalone model fit to that alone produces little of value.

Conditioning comes from a frozen self-supervised encoder — UNI for histopathology — rather than text captions, which large-image domains lack. The model descends from the same lab's earlier patch-level work and precedes PixCell, which extends UNI-conditioned diffusion to a much larger tile corpus at a single scale.

#Key Features

• Any-scale patch generation: A single checkpoint samples 256 x 256 patches at any of eight zoom levels, with the magnification supplied at inference as a conditioning input.
• Cross-magnification latent space: A trainable "summarizer" transformer compresses the variable-length array of SSL embeddings covering a patch, plus its magnification, into a fixed-size conditioning token set shared across all scales.
• Joint multi-scale sampling: Coarse-scale generations act as self-guidance for the fine scales through a linear downsampling constraint, producing 4096 x 4096 images with both local detail and global coherence in roughly 8 minutes per image.
• Reference-free synthesis: A bundled Conditioning Diffusion Model samples directly from the learned cross-magnification latent space, so generation no longer requires real embeddings extracted from an actual slide.
• Features for multiple instance learning: Intermediate U-Net denoiser activations at a fixed timestep serve as multi-scale patch descriptors for slide-level tasks.

#Technical Details

The generator is a latent diffusion model over a VQ-f4 autoencoder, with a U-Net initialized from ImageNet pretraining and cross-attention conditioning on the summarizer output; the summarizer is a 12-layer ViT-Base-style transformer. Training used 1,136 TCGA-BRCA whole slide images on three H100 GPUs at batch size 200 per GPU, yielding 12 million patches at 20x and progressively fewer at each halving of resolution. Sampling uses 50 DDIM steps with classifier-free guidance at scale 2.0. FID stays between 6.77 and 13.42 across all eight magnifications, with the largest margins over single-scale baselines at 2.5x and below, where training data is scarcest. On 4x super-resolution of 256 x 256 crops to 1024 x 1024, inferring the conditioning embeddings by inversion reaches LPIPS 0.229 on TCGA-BRCA and 0.173 on the external BACH cohort. Fusing 20x and 5x ZoomLDM features gives 94.91 AUC for BRCA subtyping and 88.03 for homologous recombination deficiency prediction under attention MIL. A parallel checkpoint trained on NAIP satellite tiles with DINOv2 conditioning demonstrates the recipe outside biology.

#Applications

The practical draw for digital pathology is synthetic imagery that holds together across the zoom range a reviewer actually uses: augmenting scarce annotated cohorts, building controlled sets for quality assurance and pathologist training, and sharing tissue-like images in place of protected patient slides. The super-resolution path targets archives scanned at low magnification, and the frozen multi-scale features offer a slide-level encoder without training a separate backbone per scale.

#Impact

ZoomLDM establishes magnification-conditioned generation as a workable route to gigapixel- scale synthesis, and its 4096 x 4096 results are produced under a compute budget that earlier coarse-to-fine and resolution-free approaches did not reach. The limits are worth stating plainly: the released pathology checkpoint covers breast tissue only, so behavior on other organs is undemonstrated, and the MIL evaluation runs on TCGA cohorts that overlap the pretraining corpus, making it a held-out-split result rather than evidence of transfer to a new cohort. Synthesis also depends on a separate frozen SSL encoder at inference. Code is MIT-licensed and the weights carry Apache-2.0 on Hugging Face; there is no hosted demo, and usage runs through the repository's inference notebooks.