Trained on 17 scribble drawings, most of them depicting celebrities.The LoRA seems to work best with Guidance Scale set at 1.5/2.0 rather than the usual 3.5. This way you get a "rougher" image, as scribble drawings usually are. Prompt following will be worse at this value, so bump it up if required. See these posts for comparisons between different Guidance Scales:<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/posts/12067470">https://civarchive.com/posts/12067470</a></li><li><a target="_blank" rel="ugc" href="https://civarchive.com/posts/12067402">https://civarchive.com/posts/12067402</a></li></ul>This LoRA compliments my crosshatch LoRA: <a target="_blank" rel="ugc" href="https://civarchive.com/models/1042061/crosshatch-drawing-style">https://civarchive.com/models/1042061/crosshatch-drawing-style</a>Known problem:Please note that sometimes an image would come out blurry because the background is white. This is a known problem with Flux, and it is not due to the LoRA (other than the fact that the LoRA, by default, will produce a white background). If that happens, you can specify "Drawing on gray paper, depicting...." so that the background will not be white. Using the beta scheduler sometimes works too.

Scribble Portrait Drawing Style

This is epoch 10, rest are at tensor. art/models/822459436858685225/Scribble-Drawing-1-1024CapD6C5-2025-01-24-21:16:19-Ep-10FLUX.1 - dev-fp8Trigger: drawingRepeat: 20 Epoch: 10 (Trained on 17 1024x1024 images for 3400 total steps)Unet LR: 0.0005 LR Scheduler: cosine Optimizer: AdamWNetwork Dim: 6 Alpha: 3<ul><li>Epoch Loss</li><li>1 0.301</li><li>2	 0.301</li><li>3 0.292</li><li>4 0.288</li><li>5 0.282</li><li>6 0.292 &lt;-</li><li>7 0.273 &lt;- lowest</li><li>8 0.283 </li><li>9 0.279</li><li>10 0.282</li></ul>

Files with SHA256: 808ed343855c5e9792ba568e1265d478b9f31502fd3b635c2c7c217970ff4dd8