There's few DMD2 upscale/refine workflows on this site but all are overcomplicated. Here's the essential simplest workflow presenting how powerful it is. I'm using this stuff at end of all image generation because its simply free 2-5sec, insane, not only upscale, but also subtle fix for every little thing. DMD2 lora - <a target="_blank" rel="ugc" href="https://huggingface.co/tianweiy/DMD2/blob/main/dmd2_sdxl_4step_lora_fp16.safetensors">https://huggingface.co/tianweiy/DMD2/blob/main/dmd2_sdxl_4step_lora_fp16.safetensors</a> At KSampler test with "start_at_step" in 60-68 range. Closer you get to 70 the less impact will it have on photo (less steps). Smaller number - more fixes (more steps) but also more intervention from the checkpoint that you potentially would not like. Input image should be in 768-1280 size. Bigger images can produce mutations like double knees.To avoid that replace "upscale image by" node with forced resize where longest side is not bigger than 2048.If you want be more aggressive with sharpness - resize image, pass it trough 1xReFocus upscaler and then do ksampler. <a target="_blank" rel="ugc" href="https://openmodeldb.info/models/1x-ReFocus-V3">https://openmodeldb.info/models/1x-ReFocus-V3</a>

Files with SHA256: 66f06ec4a3868771953c3a8e6c4425be8315dbeb02a3817d5842d524c398c49f