--Qwen2.1 Retrained with new dataset (this time took photos from s25 ultra), use with 0.7 strength--Qwen2512 <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Trained with the best my technique. Best use with NiceGirls lora (if u need better female). Also works good with other loras + works good with some nsfw loras --Qwen I have prepared a version for Qwen. It looks quite promising; however, work on its enhancement is still ongoing. Examples with girls i generated with my lora NiceGirlsP.S: yeah, I changed epoch in models manager, cause 7k steps gave overtrained result, and in process of testing I understood that 3k is enough 🗺️ BackstoryLately I kept getting DMs like “When will you drop an iPhone‑style LoRA?” Problem: iPhone LoRA already lives on Civitai—and, well, I shoot on Samsung. SamsungCam UltraReal captures the S‑series look you can’t fake with generic “phone” promptsAfter some tests i noticed that the best to use with UltraReal Fine-tune or UltraReal Abliterated with 60 steps (yeah, so much steps gives pretty good details, also text usually looks much better). Use 0.7-1.0 weight (depends on artifacts in scene)<hr />🎨 What it brings<ul><li>Vivid, high‑contrast colors</li><li>Smartphone softness – just enough in‑camera smoothing to feel “real phone,” not plastic CGI.</li><li>Highlight‑happy HDR – clipped whites around street lamps for that over‑exposed‑yet‑cinematic pop.</li><li>Less bokeh, more context</li></ul><hr />🤝 NSFWloras-friendly Tested with NippleDiffusion and MysticXXX (For some reasons is not visible in Suggested Resources) You can try to generate in 2MP for more realism and details, but be ready that in some scenes hands will be broken. But still looks very good imo in usual 1MP. In most scenes removes blurred background (sometimes it could be still blurred). Also improves details a lot even on default <a target="_blank" rel="ugc" href="http://flux.dev">flux.dev</a> (cause mostly i train all loras for my <a target="_blank" rel="ugc" href="https://civarchive.com/models/978314?modelVersionId=1413133">ultrareal fine-tune checkpoint</a> P.S: yeah, first version is slightly overtrained, but i glad with a result. v2.0 already WIP

SamsungCam UltraReal

samsung_qwen21.safetensors

Files with SHA256: d5b75fbd0a78c7642f0af49f11834d7217e3b7441ac86f2da2f598d4e0b370ee