Testing out Z-Image Base now that there's something useful to train on.It works best around 0.4-0.7, depending on the sampler and generation format. Lower weights can get really creative draws, but it's still finnicky. It does work with Z-Image turbo as well, weighted around 1.0.I've tested to confirm it composes well with other Z-Image Base LoRA's. Like most of the Z-Image testing I've done it works a lot better when there's much more detail about positioning, and the base model needs to be explicitly told to render a photograph.Basic prompt<pre><code>This is a high quality photograph of a woman sitting with her legs spread performing a footjob. Her toes are wrapped around a man's erect penis. </code></pre>More successful prompt<pre><code>photograph of a woman with long pink hair and visible tattoos on her legs, wearing a black top and partially visible patterned underwear, sitting on a bed with pillows behind her, legs spread wide, giving a footjob to a man whose erect penis is being stimulated by her toes wrapped around it, the man’s hairy legs and pubic area visible in the foreground, the woman looking directly at the camera with a neutral expression, the scene captured in natural indoor lighting with a plain light-colored wall background, high quality photograph with sharp focus on the feet and penis, composition centered on the interaction with the woman’s legs framing the action, no watermarks or digital artifacts present</code></pre>

Files with SHA256: a2a811a09b825f6ece9382b8b2889ad7b0bf0fd5e4313a8bfafb347f5ff1e879