diff --git a/README-visium.md b/README-visium.md
new file mode 100644
index 00000000..64ead82b
--- /dev/null
+++ b/README-visium.md
@@ -0,0 +1,20 @@
+# Cellxgene Visium Beta
+
+## How it works
+1. Launch `cellxgene` as normal. 
+1. If the loaded dataset has spatial information available, the image data will be loaded on startup.
+1. On the toolbar, next to the Zoom icon, a `Toggle image` button will now appear. Click on it and the image will be added as an underlay.
+1. You can now use any `cellxgene` functionality and the image will still be present. If you pan and zoom, the image will also be panned and zoomed.
+1. If you want to hide the image, you can click on `Toggle image` again
+
+In order for the image to be displayed with the correct size and alignment, the H5AD needs to have a few requirements. See the following section to learn more.
+
+## h5ad requirements
+1. The spatial embedding layer should be contained in `obsm` and be named `X_spatial`. Other layers can exist, but only this one will have the spatial feature enabled.
+2. A `spatial` dict needs to be defined in the `uns` dictionary.
+3. Inside the `spatial` dict, an `images` dict must be defined. 
+4. The `images` dict must contain a `hires` key, which should reference an image encoded as an RGB matrix (i.e., a three-dimensional matrix of size `height x width x 3` where the final dimension has the RGB values for each pixel)
+5. The `images` dict must contain a `scalefactors` dict. This should in turn contain a `tissue_hires_scalef` key, which should reference a floating point number.
+
+Moreover, in order to have the image correctly aligned with the dots, the following must be true:
+1. `tissue_hires_scalef` should represent the ratio between the embedding layer `X_spatial` and the image matrix. In particular, if you multiply `X_spatial` by `tissue_hires_scalef`, you should obtain an array of points that ovelap the tissue image if you plot them in a plane.
\ No newline at end of file