Update README.md

README.md

@@ -2,19 +2,30 @@
 license: mit
 ---
 
-# Datasets for Latent Diffusion Models for PDEs
+# Model Zoo and Datasets for Latent Diffusion Models for PDEs
 
-## Dataloading
-We use datamodules from Pytorch lightning to handle the dataset creation and dataloading. In particular, the datasets generally return a dictionary with the relevant values, controlled by flags in the config files. The basic keys that are generated are:
+## Pretrained Models
+The pretrained models are:
 ```
-- x: shape (batch, n_t, (ny), nx, c). Spatiotemporal data with variable spatial dimensions
-- pos: shape (batch, nx, d). Spatial positions of coordinates. Can include the time coordinate as well.
-- prompt: shape (batch,). A text description of x.
+- Autoencoders:
+    - ae_cylinder.ckpt : autoencoder trained to compress cylinder mesh data across 25 timesteps. Does not use GAN or LPIPS.
+    - ae_ns2D.ckpt: autoencoder trained to compress smoke buoyancy data (48x128x128). Does not use GAN or LPIPS.
+- LDMs:
+    - cylinder flow
+        - ldm_DiT_FF_cylinder.ckpt: ldm model trained to sample a cylinder flow solution conditioned on the first frame
+        - ldm_DiTSmall_FF_cylinder.ckpt: same as previous, just smaller DiT size.
+        - ldm_DiT_text_cylinder.ckpt: ldm model trained to sample a cylinder flow solution conditioned on a text prompt
+        - ldm_DiTSmall_text_cylinder.ckpt: same as previous, just smaller DiT size.
+    - ns2D
+        - ldm_DiT_FF_ns2D.ckpt: ldm model trained to sample a smoke buoyancy solution conditioned on the first frame
+        - ldm_DiTSmall_FF_ns2D.ckpt: same as previous, just smaller DiT size.
+        - ldm_DiTLarge_FF_ns2D.ckpt: same as previous, just large DiT size.
+        - ldm_DiT_text_ns2D.ckpt: ldm model trained to sample a smoke buoyancy solution conditioned on a text prompt
+        - ldm_DiTSmall_text_ns2D.ckpt: same as previous, just smaller DiT size.
+        - ldm_DiTLarge_text_ns2D.ckpt: same as previous, just large DiT size.
 ```
 
-Additionally, the datamodules store a normalizer object to scale and unscale input data. 
-
-## Cylinder Flow 
+## Cylinder Flow Dataset
 - 1000/100 train/valid samples
 - Incompressible NS in water, Re ~100-1000, dt = 0.01
 - Around 2000 mesh points, downsampled to 25 timesteps
@@ -41,7 +52,7 @@ Additionally, the datamodules store a normalizer object to scale and unscale inp
             - 'v_inlet': shape(). y-component of velocity at the inlet.
     - '1', '2', ... etc.
 ```
-## Smoke Buoyancy (NS2D)
+## Smoke Buoyancy Dataset (NS2D)
 - 2496/608 train/valid samples.
     - Datasets are divided into separates files with 32 samples each. This results in 78 training files (78x32=2496) and 19 valid files (19x32=608)
 - Smoke driven by a buoyant force, dt=1.5