Understanding LoRA Training, Part 2: Offset Noise, Epochs and Repeats
<p>Let’s talk about Offset Noise, which is supposed to be quite a crucial part of the training process.</p>
<p>Ever tried training a Stable Diffusion model against <em>a single solid black image</em>? Chances are you won’t get anything much better than this:</p>
<p><img alt="" src="https://miro.medium.com/v2/resize:fit:512/1*y_VY1PGM2WopFh3NaOLixw.png" style="height:512px; width:512px" /></p>
<p>This image comes from <a href="https://www.crosslabs.org/blog/diffusion-with-offset-noise" rel="noopener ugc nofollow" target="_blank">the key article about Offset Noise</a>.</p>
<blockquote>
<p>It seems like not only does Stable Diffusion not have the ability to produce overly dark or light images out of the box, but it also <strong>can’t even learn to do it</strong>.</p>
</blockquote>
<p>It’s a fun read!</p>
<p>If you prefer a visual explanation, here is <a href="https://www.youtube.com/@lewingtonn" rel="noopener ugc nofollow" target="_blank">koiboi</a>’s version, amazing as always:</p>
<p><a href="https://medium.com/@dreamsarereal/understanding-lora-training-part-2-offset-noise-epochs-and-repeats-c68b86c69da8"><strong>Learn More</strong></a></p>