[AI Technique] 3D -> Stable Diffusion Realism

Post Reply
xs70
Posts: 165
Contact:

Re: [AI Technique] 3D -> Stable Diffusion Realism

Post by xs70 »

Workflow, with settings for the above image: https://filebin.net/jni3lw8vgiuofyl6

You will need to have ComfyUI installed. You can drag the workflow file onto a ComfyUI window to load it.

To use the workflow:
1. Drag the image you want to retouch onto the "Load Image" node in group 2
2. Rightclick on the image -> Open in MaskEditor
3. Mask the area you want to retouch, press Save.
4. Pick values for Denoise and Steps in group 1.
5. Choose a Crop Factor in the node in group 2. The MaskDetailer node will generate a new image with dimensions equal to those of the masked area x the crop factor. Crop Factor should be a value greater than 1.0. The idea is that the node crops an image around your masked area from the original image, to provide context for the masked area, and has SD generate a new version of that image, after which the node automatically composits the masked area from the new image into the original.
6. Hit Queue in ComfyUI, repeat until you like the result. Repeat for every subject in the image until all is retouched.
Last edited by xs70 on Wed Oct 08, 2025 6:24 am, edited 5 times in total.
xs70
Posts: 165
Contact:

Re: [AI Technique] 3D -> Stable Diffusion Realism

Post by xs70 »

The model is Cyberdelia's Realistic Illustrious model: https://civitai.com/models/1125067?mode ... Id=1973422

Cyberdelia specializes in realistic versions of SD models (SDXL, Pony, Illustrious, Flux). I use the Illustrious version, because Illustrious works with Danbooru tags and understands some GIMPy concepts: https://danbooru.donmai.us/posts?tags=peril&z=5. Version 5.0 has the highest level of realism (for now).
xs70
Posts: 165
Contact:

Re: [AI Technique] 3D -> Stable Diffusion Realism

Post by xs70 »

masked.jpg

Screenshot from inside the ComfyUI mask editor showing the masked area that I used for the image above (in white). Note that you don't need to be precise, it's often better to provide some extra padding in neutral areas surrounding your subject (like the floor or the wall behind her in this image) so the MaskEditor node can blend/feather the new result into the original image.
Last edited by xs70 on Wed Oct 08, 2025 11:55 am, edited 1 time in total.
xs70
Posts: 165
Contact:

Re: [AI Technique] 3D -> Stable Diffusion Realism

Post by xs70 »

PS:
doe.1971 wrote: Mon Oct 06, 2025 8:34 pm I was wondering if you know any realistic XXX model for Comfy at 12 or 16 GB even better
I'm currently experimenting with WAN 2.2 (14GB) as a realism finisher. WAN is trained on video as well as images, which enhances the realism, and the T2V models can also be used to create/inpaint still images. A strong candidate for a finisher model. Early days though, I only got it working yesterday.

That said, I think you'll be able to do a lot with CyberIllustrious v5.
User avatar
doe.1971
Posts: 1872
Contact:

Re: [AI Technique] 3D -> Stable Diffusion Realism

Post by doe.1971 »

Thanks boss, I'll try it in a day or two. Now I have the main PC opened to upgrade the power source, the new card addition is sucking some 300W compared to 100W on the old one and the old PSU was dangerously close to the rated power capacity. Add a 16 cores Ryzen 9 and you need a small modular reactor. :)

Frankly I don't plan to use AI in any BDSM creative capacity, I'd rather do it in Poser, get more control even if it doesn't looks so great. What I need is to correct those human joints that are looking odd, due to the extreme positions I put my models into. Or some chromatic corrections, small things like these. Or some skin that looks weird because of lighting issues.
Post Reply

Return to “Art Contributions”

Who is online

Users browsing this forum: Goijeraboy, shankster and 21 guests