MONET (Morphological Observation Neural Enhancement Tool) is a diffusion model that generates virtual cell paint images from brightfield microscopy. Cell painting is a popular technique for creating high-contrast images of cell morphology, but it is labor-intensive and requires chemical fixation—making time-lapse studies impossible. MONET bypasses these limitations by predicting five cell paint channels (DNA, RNA, ER, cytoskeleton/AGP, and mitochondria) directly from brightfield images.
The model uses a reference consistency architecture that enables artifact-free generation of time-lapse videos, despite the fact that paired (brightfield, cell paint) video training data cannot exist. At inference time, the first frame is generated unconditionally, then subsequent frames are conditioned on that first frame to maintain visual consistency. This architecture also enables a form of in-context learning for domain adaptation to new cell lines and imaging hardware.
MONET is a 350M parameter UNet-based diffusion model trained on 8M+ images from the Broad Cell Paint Gallery. Pre-trained weights are available on HuggingFace at
IntegratedBiosciences/monet.
