<a target="_blank" rel="ugc" href="https://patreon.com/psoft"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8dcbc56a-15e9-429e-9b36-c2bc22594678/width=525/8dcbc56a-15e9-429e-9b36-c2bc22594678.jpeg" /></a> Support my work on <a target="_blank" rel="ugc" href="https://www.patreon.com/psoft">Patreon</a> and <a target="_blank" rel="ugc" href="https://ko-fi.com/psoft">Ko-Fi</a> and get access to tutorials and exclusive models. Rate and leave a like if you enjoyed it, and follow for new models!<hr />LoRA trained to generate a person in public surrounded by people, such as photographers. Works with other LoRAs and various models.UPDATE - new training and small improvementsUpdated training: used LyCORIS instead of LoRA, trained on larger dataset with more realistic images, cleaned up tags.Updated output: less likely to have a camera on character's hands, more detailed crowd, a bit more consistent, you can better increase the crowd or photographers by increasing the trigger tag's weight.Main tags (use one or both for different effect): public, photographers// OLD VERSION (1.0) //Main trigger word: <code>public </code>Other tags used <code>photographer</code>,<code>phone screen</code>,<code>camera</code>,<code>foreground</code>You can use your usual CFG and weight <code>1.0</code>See weight comparison 0.0 vs 1.0, minimal effect on style and composition:<img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e9f913bc-38d5-47c1-25fc-c826934c9600/width=525/e9f913bc-38d5-47c1-25fc-c826934c9600" /><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dfba320d-9bf6-4350-7d82-9bdbcf499300/width=525/dfba320d-9bf6-4350-7d82-9bdbcf499300" /><img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/47814eb6-e6a1-4e97-fc5a-342145ba6d00/width=525/47814eb6-e6a1-4e97-fc5a-342145ba6d00" />

In public / photographers

Trained on NAI using kohya GUI218 images (unedited, not cropped), auto tagged with WD1.4, added "public" token, "keep token = 1" and shuffle tokens.218 images * 10 repeats * 5 epochs / batch 6LR 0.0002, AdamW8bit, constant512,512 resolution, bf16

Files with SHA256: d18d38b05f536718a256b4721b873c6095935f5abc7c62ef1c55835c49d220d7