Simply put, this is a ComfyUI workflow to batch image-to-image process rendered scenes from something like a Ren'py game with DAZ3D-rendered graphics to photo-real scenes. It uses the following technique: <ul><li>WAS Nodes to batch load and save images in a directory to another directory. Use either Auto-queue or set the batch number to the amount of images you wish to process. Keep in mind that WAS does not order filenames the same way windows does.</li><li>WD14 Tagger to trigger certain Loras to load when certain things are included in the source image, eg. loading a blowjob-related Lora when WD14 detects the word "fellatio". It also switches to a lower Denoise value when "close-up" is detected.</li><li>Conditioning (Set Mask) to try to make the men and women look more like their respective genders. Using the condition "redhead woman" also kind of works so that you can apply an embedding directly to the area it detects, but this only kind of works.</li><li>3 Controlnets: softedge (HED lines), depth map (Zoe), and mediapipe face to help preserve the scene composition.</li><li>UltimateSDUpscale as the primary I2I sampler - for screen aspect ratio, this seems to work best when I downscale to 1MP so that there's 2 768x768 tiles, then upscale.</li><li>FaceDetailer x 2 - once for the faces, once again for the genitalia.</li></ul>Custom Nodes used: WD14 Tagger, Impact Pack, CLIPSeg, FaceDetailer, Auxillary Controlnets, WAS Nodes, UltimateSDUpscale, maybe some others??UPDATE 1.1: I disabled the 2nd Detailer and one of the controlnets, and also the 3rd conditioning mask. This version does two passes - like a low-res and a high-res (at the full output resolution), which I think yields better results, at the cost of speed.

Files with SHA256: cef99bde4b3ffd6694b25f22e0928a32d06f15ad29676a6ada70e0da443ba785