<p>This model is trained on a dataset composed almost entirely of real-world photography. The primary goal was to achieve realistic textures and lighting, moving away from the "plastic" look common in purely synthetic datasets. While the model is an ongoing project, it performs reliably for its intended purpose.</p><p>This version is based on Krea 2, which produces highly stylistic results. It works well with weights set to 1 and handles uncensored content effectively. (Note: I apologize for the delay with this release, as I was quite slow to publish a LoRA that had already been on my computer for nearly a month).</p><hr /><h3 id="key-features-and-usage">✨ Key Features &amp; Usage</h3><ul><li><p><strong>No Hard Trigger Word:</strong> You generally do not need a specific trigger word. If you want to emphasize size, adding keywords like "huge" or "big" will prompt the model accordingly.</p></li><li><p><strong>Body Type Adjustments:</strong> Training on real-world data often biases models toward generating curvier body types. This model attempts to maintain a balanced default; however, if you prefer a slimmer physique, adding "skinny" to your prompt should adjust the body type without issues.</p></li><li><p><strong>Wardrobe &amp; Content Notes:</strong> The model handles nudity well. Please note that if your prompt includes "cleavage," there is a probability of generating an areola slip. If this is not the desired outcome, generating with a different seed or lowering the model weight may help.</p></li></ul><hr /><h3 id="recommended-workflow">⚙️ Recommended Workflow</h3><p>For highly detailed and realistic images, the following setup is recommended:</p><ol><li><p><strong>Base Model:</strong> Use <strong>'Krea 2 Raw'</strong> with a speed LoRA applied. This configuration tends to produce more varied and detailed results, giving you the best images. You can drag and drop the sample images into ComfyUI to load the exact workflow used.</p></li><li><p><strong>Filter Bypass (Optional):</strong> An additional LoRA (if it can even be called a LoRA, given the complex engineering behind it) created by <a rel="ugc" href="https://civarchive.com/models/2728234/krea2filterbypass">S1LV3RC01N</a> was used for the cover images. I want to clarify that the MRB model of Krea 2 works well even without the filter bypass. However, I highly recommend using it, as it spice up a little the images and also seems to help with prompt coherence.</p></li></ol><hr /><h3 id="current-limitations-and-future-plans">⚠️ Current Limitations &amp; Future Plans</h3><ul><li><p><strong>Future Updates:</strong> The next version of this model will take some time, as I am currently focused on improving the Ideogram version. I hope to implement a small project I have in mind; if it succeeds, future versions of my models will be developed using Reinforcement Learning from Human Feedback (RLHF). I hope to achieve this to gain better control over the training process and get the models to perform exactly as intended.</p></li></ul><hr /><h3 id="support">🧡 Support</h3><p>if you want to support my work, you can contribute here:<br /><a target="_blank" rel="ugc" href="https://ko-fi.com/flusk">https://ko-fi.com/flusk</a></p>

Files with SHA256: 68654fe7342a21cfa3bcd35f036d5f0a77f3d9c2b1c74e9eec308f15c0cb5c83