<p><strong><span style="color:#228be6">Please share your creations in the model gallery. Thank you.</span></strong></p><p>This Lora generates drawings in the style of Charles Dana Gibson. Gibson was a very popular artist at the turn of the 20th century. His depictions of women were so popular that the term "Gibson girl" entered the vernacular as a description of women who looked liked the idealized characters in his sketches.</p><h1 id="heading-5262">Versions:</h1><h2 id="heading-5263">Gibson XL - <span style="color:rgb(250, 82, 82)">New!</span></h2><p>New SDXL version. A little less expressive, a little more coherent than SD1.5 versions.</p><h2 id="heading-5263">cdg-v1</h2><p>cdg_v1was the first iteration. Trained at 768 resolution and a low learning rate. This model was trained on a dirty image set which included captions, text, signatures and watermarks. Every image was weighted equally in the training. It still works great and is pretty flexible.</p><h2 id="heading-5264">cdg_v2</h2><p>cdg_v2 was trained on a cleaned up image set without captions, signatures, watermarks or text. Images were weighted by quality, resolution, and differentiation from contemporary illustrators. Also trained at 768 resolution and a low learning rate. This version is pretty precise in capturing the Charles Dana Gibson style. This version was also trained using WD14 captions, so captions are parsed, making the model more flexible in styling atypical scenes.</p><h2 id="heading-5265">cdg_v3</h2><p>cdg_v3 was trained on the improved and weighted dataset at 512 resolution and a higher learning rate. The training set used BLIP captions of whole sentences. Not sure what impact this will have on inferences created with it. I think it creates the best pictures.</p><h2 id="heading-5266">Trigger Words</h2><p>All of them can be used with or without trigger words to different effect. Using the trigger words delivers results that are more orthodox to Charles Dana Gibson's style. Sometimes, at higher CFGs, using trigger words will break the images. Try with or without.</p><h2 id="heading-5267">Clip Skip</h2><p>Can use at Clip skip 1 or 2 to different effect.</p><h2 id="heading-5268">CFG</h2><p>Can use CFG 1-15, to different effect. Use appropriate steps to keep images from breaking.</p><h2 id="heading-5269">Weight</h2><p>Generally, use at full strength of 1.0. Sometimes need to bump up to 1.2 or more to overcome the strengths of additional Lora used in an inference.</p>

Charles Dana Gibson XL/SD15

<p>Cleaned up the dataset in this version, getting rid of signatures, page yellowing, text and captions.  Trained at 512 x 512 resolution and a higher learning rate. Used BLIP captions.  It can be used with trigger words or without trigger words to different effect.  Use "drawing by Charles Dana Gibson" or "drawing by g1bs0n" or both or neither. Best at full strength.</p>

Files with SHA256: 644e34b854987038590aa513992a4a4c2374303de8311b8cfa355a0abc0efe17