This tool lets you find images that are similar to a query — either an image you upload, or a short text description — from SSBD's shared OME-Zarr imaging datasets. It compares embedding vectors (numeric fingerprints produced by a pretrained vision model), not the raw pixels directly.
File requirements:
What happens to your image before comparison:
It's converted to grayscale first — the SSBD-side images are grayscale scientific images, so discarding color means the comparison is driven by structure and texture rather than color. What happens after that depends on the model you picked:
The SSBD-side images go through their own, different preparation before reaching this same per-model step: each dataset is multi-dimensional (multiple channels, Z-slices, and sometimes timepoints), so it's first reduced to one 2D grayscale plane per channel (a maximum-intensity projection across Z, at one representative timepoint), with pixel intensities contrast-stretched to a normal image range.
Describe what you're looking for in a sentence or a few words (e.g. "meiosis", "mouse", "neuron"). Text search only works with CLIP, since it is the only one of the three models with a text encoder — DINOv2 and uniDINO are vision-only and have no way to represent text.
Because the corpus images are grayscale and quite different from the natural photos CLIP was mostly trained on, text search quality varies: conceptual terms (e.g. process names) tend to work better than organism names, and CLIP has no notion of physical scale, so structurally similar things at very different physical scales can be confused.
Both search modes let you pick which embedding model to use:
They were trained on different kinds of data with different objectives, so the same query can rank matches quite differently depending on which one you pick:
Matched channel & all channels:
Your uploaded image is a single flat picture — it is not split into channels. SSBD's data, however, is often multi-channel fluorescence microscopy. Rather than merging a dataset's channels into one embedding, this tool computes a separate embedding for each channel independently and compares your query against all of them. That's why each result shows which specific channel matched ("matched channel"), alongside an "all channels" composite image (all channels merged with their original display colors) so you can see the matched channel in the context of the whole image.
Similarity score:
Each embedding is a unit vector, and the similarity score is the cosine similarity between your query's vector and a candidate's vector — in practice this is just their dot product, since both are already unit length. It ranges from −1 to 1, where 1 means the two vectors point in exactly the same direction (as similar as that model can represent) and 0 means they are unrelated by that model's measure. Scores are only comparable within the same model — CLIP, DINOv2, and uniDINO each have their own scale and typical range, so a 0.4 from one is not equivalent to a 0.4 from another.
Similarity threshold:
Hides candidates scoring below a chosen cutoff (default 0.3). You can change it and re-apply without re-running the search — it just re-filters the same ranked results, so it's free to experiment with.
Important — similarity is not a biological match:
A high score means the two images look alike to that particular pretrained model — it is not a statement about the underlying biology. None of these three models were trained or validated to recognize organism identity, cell type, experimental condition, or any other biological label. Visually similar structures can come from completely different organisms or contexts (for example, a search for one organism's images can surface a different organism's images that simply share similar-looking cellular structures), and conversely, biologically related images can score low if they don't happen to look alike to the model. Treat results as a starting point for visual exploration, not as an identification or classification.
The original file you upload is not stored. What is kept, temporarily, is a small (up to 200×200px) preview thumbnail and the embedding vector computed from it — just enough to show "Your query image" at the top of the results and to let you page through results without re-uploading. This is kept in your browser session (server-side, tied to your session cookie), not shared with other visitors.
It is replaced as soon as you run a new search (image or text), and otherwise expires automatically within 1 hour, or immediately when you close your browser — whichever comes first.
ViT-B-32-quickgelu model with OpenAI's original weights).
facebook/dinov2-small checkpoint).