lora (weight) from -0.5 to 1minus - is more cozy / natural / bit more bokeh plus - is more straight mostly more background, less bokeh, bit less face-detailslook -&gt; image postbokeh have two versions (lora (weak / strong) from -0.4 to 0.8you can use CFG-Scale higher than 2, it generates slower i know, but it is the requirement for mor details like the description on FLUX_Distilled <a target="_blank" rel="ugc" href="https://civarchive.com/models/843551?modelVersionId=943787">https://civarchive.com/models/843551?modelVersionId=943787</a> training were pictures with patterns and silhouettes, including the inversion (negative image) without captions.it adds details, it adds a background at a simple prompt even if you don't call it.<hr />What i have found out that . . .first the schedule-type is important (DDIM or SGM-Uniform or BETA) you can also try FluxRealisitic with this three schedulers<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c0f27d27-df61-4e79-82f2-692f2c8e5454/width=525/c0f27d27-df61-4e79-82f2-692f2c8e5454.jpeg" />second the prompt order and some usefull trigger are importanteg:"detailed, sharp outlines, contrast, wide angle lens captures the entire scen with a street or clouds (your words for details in background eg: car, people, building)",and then the object of interest"a woman in front posing on a street"<hr />and on simple prompt eg: "photo of a woman" it adds at least a blurry background<hr />avoid the word photo, portrait that makes more bokeh<hr /><hr />play with it ... and tell me

ANTI Bokeh - Add detail - CFG

same image set, new train-parameteri have linked an image post

Files with SHA256: 0fc8ff31ead7764e8f336bb5d06d950e6a47be379ed1b219e36f897338e47b3e