<h1 id="exciter_hr-redemption">Exciter_HR - Redemption</h1>Take the int8 version, the bf16 versions is only for people with the black output problem. it does not provide a better quality, it will use double the VRAM and double the generation time.<h3 id="new-in-redemption-release:">New in Redemption release:</h3><ul><li>Even higher image quality and general aesthetics</li><li>Better lighting, more color depth (higher dynamic range), better macro structures.</li><li>Perfect smooth gradations most of the time (I love them so much!), zero blotches even at int8</li><li>Even faster. Perfection at 7steps with euler_ancestral_pp/simple</li><li>Even more unique faces per seed and generation</li></ul><h3 id="what-do-i-need-to-know-about-the-redemption-release:">What do I need to know about the Redemption release:</h3><ul><li>Text needs more steps, try 9.</li><li>Images have more depth of field (cinematic), if you want sharp backgrounds, use (detailed background:2.0) and (Depth of field, DOF:-1) in your prompt. "Krea2 Prompt Weight" Node must be used.</li><li>The new method used for acceleration can cause body horror or strange artifacts in very rare cases (under 1%).</li><li>Between this and the original release are 11 versions of the model that did not deliver a better generation, therefore iteration 20 is a "Redemption" for me.</li><li>The model is very different from the initial release, so more like an alternative then a better one. Ok, it is far better, - but I guess that’s all subjective \o/.</li></ul>I am working on one even better version, and hopefully it will not fail another 10 times. Leave a like if you want me to continue, I need some motivation :0)<h1 id="krea2-turbo-exciter_hr">Krea2 Turbo - Exciter_HR</h1>Take the int8 version, the bf16 versions is only for people with the black output problem. it does not provide a better quality, it will use double the VRAM and double the generation time.<h3 id="what-can-it-do">What can it do?</h3><ul><li>Has turbo capacities</li><li>Has 90% more variation in generations across different seeds thanks to secret sauce</li><li>Is fully uncensored with zero image degradation</li><li>Every generated face is unique</li><li>Is trained on 15000 highest quality Hi-Res Images</li><li>90% reduced artifacts (moles and blotches) thanks to secret sauce</li><li>Does direct generations up to 3840 x 2160 (in only 3-4 minutes and 12GB VRAM)</li><li>Works with loras</li><li>Works with Krea2 Prompts weight node</li><li>Works with NegPip node</li><li>Works with cfg 1.0 - 1.2 and negative clip prompting</li><li>Works with very simple prompts and delivers a wide variety of outcomes</li><li>Works with complex prompts (e.g. prompt enhancer)</li><li>It still can do anime, illustrations, oil paintings, comics and celebs etc.</li></ul><h3 id="ok-but-at-what-is-it-bad">Ok, but at what is it bad?</h3><ul><li>Does sometimes body horror (2-3%)</li><li>Applies sometimes the wrong gender genitals</li><li>Has sometimes problems with thin line patterns</li><li>Man does it suck with text mostly</li><li>Does sometimes add a corner logo or web addresses or watermarks (under 2%)</li><li>When prompted contradictory, It will do it without self-correcting (expect body horror)</li><li>Simple natural prompts of sex positions will probably not work, use the prompt enhancer, or learn how to formulate them correctly.</li><li>It can corrupt you</li></ul><h3 id="anything-i-need-to-know">Anything I need to know?</h3><ul><li>Be careful what you prompt for, the model will deliver relentlessly.</li><li>Stacked loras with accumulated weights over 2.0 will re-introduce the famous blotchy artifacts.</li><li>works well with:</li><li>euler/simple/12steps (natural, prone to macro errors)</li><li>heunpp/simple/6steps (softer image)</li><li>exp_heun_2_x0_sde/simple/6 steps (super clear image)</li><li>euler/sigma_curve_from_point/12steps (fast, super clear image, strong colors)</li><li>euler/beta_57/9steps (sharp, clean, fast)</li><li>euler_ancestral_cfg_pp/simple/12steps (godlike quality, slow)</li><li>Use this Text Encoder and no other, especially for erotic generations with text enhancer:</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Zynerji/Ektome-Qwen3-VL-4Bi-PristinelyUncensored">https://huggingface.co/Zynerji/Ektome-Qwen3-VL-4Bi-PristinelyUncensored</a></li><li>Use this VAE for the best image quality:</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2718533/hdr-vae-anima-krea2-qwen-image">https://civarchive.com/models/2718533/hdr-vae-anima-krea2-qwen-image</a></li><li>Use this simple workflow if you want:</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2856123/krea2turbo-exciterhrworkflow?modelVersionId=3225778">https://civarchive.com/models/2856123/krea2turbo-exciterhrworkflow?modelVersionId=3225778</a></li></ul><h3 id="prompting-tips:">Prompting tips:</h3><ul><li>Starting the prompt with "real photo of" makes sure the image looks natural and not 1girl like or like a CGI render.</li><li>If you prefer large breasts or wide hips over a natural body you have to prompt for it.</li><li>For females use the word woman, - do not use the word girl (for gods sake).</li><li>For fun try just one word like fox or bee to see what the model can do with a different seed. (because it can)</li></ul><h3 id="additional-info">Additional Info</h3>The version released here, is not the first version, it is the 10th iteration of the model being trained for the last 2 months.The training data contains a wide variety of EXCLUSIVE images, from buildings to animals and nature and people with over 2000 different poses and body positions incl. erotic.The model has full knowledge of human anatomy (Inc. pubic hair, breast sizes, body shapes and muscles). No wd14 captions are used. All captions are full text and every single one has been checked and corrected by human eyes (yes mine, and it almost made me insane) to make sure it's not wrong.And now load it and have some fun!

Files with SHA256: ab63bc267dbf55af8bba5ebe92706a08ea947edd66a5e404f1fe2a97b30d837f