<h1 id="heading-16034">Karsh Style XL</h1>A funny thing about one of the most prolific portrait photographers to have graced the 20th century—I am certain that anybody today has seen an influential person of that era through his camera, yet the name of the man behind the lens is lost upon them.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e2d214cd-be60-41d2-9673-544b85dacc00/width=525/e2d214cd-be60-41d2-9673-544b85dacc00.jpeg" />Yousuf Karsh was a photographer who is remembered for his talent in portraiture, and an impressive list of clientèle that spanned many influential and powerful individuals from the past century. He had an exceptional ability to control lighting in his compositions, and an uncanny gift in conditioning his subjects to appear dignified and heroic, regardless of their vocation or position in society.Karsh often crosses my mind whenever I think of dramatic lighting or scene composition, and I find myself sometimes attempting to replicate his style. I hope that this LoRA helps others who wish to do the same.<h2 id="heading-16035">tl;dr</h2><ul><li>Juggernaut XL v7 works well.</li><li>Positive prompt:<ul><li>B&amp;W: monochrome karsh style, &lt;subject prompt&gt;, &lt;additional hints&gt;, {close|medium|full|wide} shot</li><li>Colour (untested): karsh style, &lt;subject prompt&gt;, &lt;additional hints&gt;, {close|medium|full|wide} shot</li></ul></li><li>Negative prompt:<ul><li>Use your favourite base prompt.</li><li>(untested) Add "monochrome" to the negative prompt if colour images are desired.</li></ul></li></ul><h2 id="heading-16035">Usage notes</h2><ul><li>This model will help SDXL-based checkpoints draw with qualities reminiscent of Yousuf Karsh's photographs, including imparting similar:<ul><li>lighting style and tonality.</li><li>scene compositions and details in subjects.</li></ul></li><li>Checkpoints:<ul><li>Any checkpoint capable of generating photorealistic images should work.<ul><li>Juggernaut XL v7 works well.</li></ul></li><li>This model can generalize to illustrations, but it is difficult to yield good lighting.<ul><li>kohakuXL-beta v7 can work (for fun).</li></ul></li></ul></li><li>Positive prompt:<ul><li>This prompt format appears to work:<ul><li>[monochrome ]karsh style, &lt;subject prompt&gt;, &lt;additional hints&gt;, {close|medium|full|wide} shot</li></ul></li><li>Feel free to experiment. The three aspects of the suggested prompt format that are grounded in training are:<ul><li>"monochrome", which can be removed to draw colour images</li><li>"karsh style", which is the activation phrase</li><li>"{close|medium|full|wide} shot", which are zoom/crop controls that have been tagged on each image in the dataset.</li></ul></li></ul></li><li>Negative prompt:<ul><li>Use your favourite base prompt.</li><li>Add "monochrome" to the negative prompt if:<ul><li>colour images are desired.</li><li>the model does not yield colour after removing "monochrome" from the positive prompt.</li></ul></li></ul></li><li>Art style notes:<ul><li>I have performed little testing on the full characteristics of this model, and instead only focused on black and white portraiture.</li><li>The model appears to:<ul><li>subtlety modify the lighting and tonality characteristics of a scene.</li><li>retain most subjects in the scene, but modifies objects to be less cluttered and people to carry more dignified expressions.</li></ul></li><li>Landscape and candid street photography images were included in the training dataset, but not tested for recall.</li><li>Colour images were included in the training dataset, but initial evaluations did not impress as much as black and white. Further testing is needed here.</li><li>There appears to be some gender/race/age bias from the dataset, but they are easily overridden by positive prompting for desired characteristics.</li></ul></li><li>Have fun!</li></ul><h2 id="heading-510">Release notes</h2>v1.0 (2023-12-27)<ul><li>Initial release.</li></ul><h2 id="heading-16036">Model information</h2>This model was trained against photographs taken by Yousuf Karsh. Art was selected by hand. The dataset was largely constrained by low availability of street/landscape and colour photographs against black and white portraiture.See the "About this version" card for information about the dataset and training parameters.<h2 id="heading-945">Public domain notice and disclaimers</h2>The author of this generative AI model (model) releases rights and disclaims liabilities that pertain to the model itself or usage thereof, under the mark of the CC0 1.0 legal code.<a target="_blank" rel="ugc" href="http://creativecommons.org/publicdomain/zero/1.0">http://creativecommons.org/publicdomain/zero/1.0</a>This work derives from public domain and copyrighted photographs that were created by Yousuf Karsh in his lifetime. It is not endorsed by his estate, nor is it intended to generate images that imply his authorship.

Karsh Style XL

Training run 1, epoch 8Dataset 1Images<ul><li>Count: 59 images</li><li>Sources:<ul><li>karsh.org</li></ul></li><li>Treatment:<ul><li>All source images were used as-is.</li></ul></li><li>Flip augmentation: enabled</li></ul>Tagging<ul><li>Methodology: Images were initially tagged by BLIP (model_base_caption_capfilt_large). Tags were edited by model author for correctness and consistency.</li><li>Tags were structured following this pattern:<ul><li>Tag 1: "[monochrome ]photograph in karsh style of"<ul><li>"monochrome" omitted if picture has colour.</li></ul></li><li>Tag 2: a brief description of the subject and scene</li><li>Tag 3: "{close|medium|full|wide} shot"<ul><li>"close", "medium", "full" largely used for portraiture.</li><li>"wide" largely used for landscape/street photographs, or photographs where a person was not the primary focus.</li></ul></li></ul></li><li>Tag order was preserved during training.</li></ul>Training<ul><li>Dataset replicates/epoch: 10</li><li>Epochs: 8</li><li>Batch size: 4</li><li>Checkpoint: sd_xl_base_1.0.safetensors</li><li>Network configuration: Kohya Standard (dim=16, alpha=8)</li><li>LR: 0.0001 for both U-Net and text encoder. Algorithm was "cosine_with_restarts" with 5% warmup.</li><li>Optimizer: AdamW</li></ul>

Files with SHA256: 87d689d9ad21d2fc144492a4a75bc812847a5ffb52ff9f1d603af463a1542d59