<h1 id="uhdman-krea-2-nsfw-capable-&quot;edit&quot;-model">UHDman - Krea 2 NSFW Capable "Edit" Model</h1><h3 id="uhdman-is-my-attempt-at-training-an-adonis-upscale-model-for-krea-2.">UHDman is my attempt at training an Adonis upscale model for Krea 2.</h3>This was not a success in that area, as the images are more of a loose reinterpretation vs an edit like in Flux 2 Klein 9B.<h2 id="this-model-instead-does-the-following:">This model instead does the following:</h2><ul><li>Adds further male anatomy knowledge to Krea 2 for enhanced realism</li><li>Enhances both male and female bodies/textures for SFW and NSFW generations</li><li>(Mostly) Eliminates the need for a bypass lora to generate NSFW imagery</li><li>Generates at high resolution (3.0MP recommended), the model was trained on 1024, 1280, and 2048 target resolution images with varied control image sizes like the Adonis models</li><li>It will follow basic editing directions, the input image needs to be as close as possible to the wanted output, it behaves like creative image-to-image vs a real edit model but things like identity and lighting can usually be changed.</li><li>The model can be used for high quality reinterpretations of SFW and NSFW images. Like the Adonis models, training on high resolution images of skin and bodies helps with texture all types of generations.</li><li>May work as a regular lora but I haven't tested it much for that use case. </li></ul>The trigger word is uhdmanscale. The dataset used the same captions from the Klein 9B dataset (which helped with avoiding training unwanted concepts) but the used captions have little effect towards steering the model one way or another. An input image and just "uhdmanscale" as a prompt will generate a similar-ish image, with some basic prompt changes usually followed to create a "new" image.<h3 id="using-cfg-1.5-to-2.0-greatly-improves-detail-with-a-small-speed-penalty-vs-cfg-1">Using CFG 1.5 to 2.0 greatly improves detail with a small speed penalty vs CFG 1</h3>You will need the following models:<a target="_blank" rel="ugc" href="https://huggingface.co/Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8/blob/main/Krea2_Raw_convrot_int8mixed.safetensors">Winnougan/Krea2_Raw_convrot_int8mixed.safetensors (Or another version of Krea2_Raw)</a><a target="_blank" rel="ugc" href="https://huggingface.co/Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8/blob/main/Text%20Encoder/qwen3vl_4b_uncensored_int8_convrot.safetensors">Winnougan/qwen3vl_4b_uncensored_int8_convrot.safetensors (Or another version of qwen3vl_4b)</a><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Krea-2/resolve/main/loras/krea2_turbo_lora_rank_64_bf16.safetensors?download=true">Comfy-Org/krea2_turbo_lora_rank_64_bf16</a><a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Wan2_1_VAE_fp32.safetensors?download=true">Kijai/Wan2_1_VAE_fp32</a><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/flux2-dev/blob/main/split_files/vae/flux2-vae.safetensors">Comfy-Org/flux2-vae.safetensors (For final output VAE decoding)</a>For the workflow you will need the following nodes:<a target="_blank" rel="ugc" href="https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit.git">Krea2-Ostris-Edit</a><a target="_blank" rel="ugc" href="https://github.com/rgthree/rgthree-comfy">rgthree</a><a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KJNodes">KJNodes</a><a target="_blank" rel="ugc" href="https://github.com/ClownsharkBatwing/RES4LYF">RES4LYF</a><a target="_blank" rel="ugc" href="https://github.com/vrgamegirl19/comfyui-vrgamedevgirl">vrgamedevgirl</a><a target="_blank" rel="ugc" href="https://github.com/BigStationW/ComfyUi-Scale-Image-to-Total-Pixels-Advanced">ComfyUi-Scale-Image-to-Total-Pixels-Advanced</a>

UHDman - Krea 2 NSFW Capable "Edit" Model

v1 of this model instead does the following:<ul><li>Adds further male anatomy knowledge to Krea 2 for enhanced realism</li><li>(Mostly) Eliminates the need for a bypass lora to generate NSFW imagery</li><li>Generates at high resolution (3.0MP recommended), the model was trained on 1024, 1280, and 2048 target resolution images with varied control image sizes like the Adonis models</li><li>It will follow basic editing directions, the input image needs to be as close as possible to the wanted output, it behaves like creative image-to-image vs a real edit model but things like identity and lighting can usually be changed.</li></ul>

Files with SHA256: 578376e8acbc00c35bb81c56d066b4ade1ba527cbfb69ab3ba0973e763773ce9