Updated project about page link.

_layouts/about.liquid

@@ -54,6 +54,7 @@ layout: default
         <a href="{{ '/blog/' | relative_url }}" style="color: inherit">Latest posts</a>
       </h2>
       {% include latest_posts.liquid %}
+      {% include latest_projects.liquid %}
     {% endif %}
 
     <!-- Selected papers -->

_projects/2024-05-26-fine-tune-stable-diffusion-vae.md

@@ -21,7 +21,7 @@ scholar:
     bibliography: references.bib
 ---
 
-In a recent project, we (my coauthor and myself) needed to train a denoising diffusion bridge model on 512x512x3 patches taken from 2048x2048 fundus images of human eyes. As GPU memory requirements for training a diffusion model on high resolutions with a U-NET backbone is prohibitive, scaling quadratically with image resolution, we turned to latent space diffusion models. In particular, we wished to use a fine-tuned version of the auto-encoder from Stable Diffusion. Unfortunately, it was somewhat of a lengthy process finding exactly what training parameters worked well for fine-tuning Stable Diffusions VAE, or a short guild with training scripts. To address this, I have written this short article and accompanying [github](https://github.com/SpencerSzabados/Fine-tune-Stable-Diffusion-VAE) repo, which is based on material from [capecape](https://wandb.ai/capecape/ddpm_clouds/reports/Using-Stable-Diffusion-VAE-to-encode-satellite-images--VmlldzozNDA2OTgx) and [cccntu](https://github.com/cccntu/fine-tune-models).
+In a recent project, we (my coauthor and myself) needed to train a denoising diffusion bridge model on 512x512x3 patches taken from 2048x2048 fundus images of human eyes. As GPU memory requirements for training a diffusion model such high resolutions with a U-NET backbone is prohibitive, scaling quadratically with image resolution, we turned to latent space diffusion models. In particular, we wished to use a fine-tuned version of the auto-encoder from Stable Diffusion. Unfortunately, it was somewhat of a lengthy process finding exactly what training parameters worked well for fine-tuning Stable Diffusions VAE, or a short guild with training scripts. To address this, I have written this short article and accompanying [github](https://github.com/SpencerSzabados/Fine-tune-Stable-Diffusion-VAE) repo, which is based on material from [capecape](https://wandb.ai/capecape/ddpm_clouds/reports/Using-Stable-Diffusion-VAE-to-encode-satellite-images--VmlldzozNDA2OTgx) and [cccntu](https://github.com/cccntu/fine-tune-models).
 
 ---