- Jul 12, 2017
- 1,154
- 1,104
Any suggestions on how to make a FFM/MMF combo without the female somehow have two different features on the same body?
- Jun 28, 2017
- 312
- 504
Honestly these are quite good - I'm getting dofantasy vibes
Mini story I made with a series of images:
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
- Jun 9, 2019
- 344
- 1,335
For those who use forgeUI, do try to use SelfAttentionGuidance option (You can download it as a extension on automatic1111). It makes the picture more detailled.
I use this setting
I suggest trying to tweak it to get what you want. Don't go too high or the picture is going to be saturated, and too low, no change will happen.
Here's an exemple:
With SelfAttentionGuidance
Without SelfAttentionGuidance
Open both pics in full view, each picture on its own tab, then switch back and forth to see the difference.
I use this setting
You don't have permission to view the spoiler content.
Log in or register now.
I suggest trying to tweak it to get what you want. Don't go too high or the picture is going to be saturated, and too low, no change will happen.
Here's an exemple:
With SelfAttentionGuidance
Without SelfAttentionGuidance
Open both pics in full view, each picture on its own tab, then switch back and forth to see the difference.
Good news: i trained new LoRA.
Bad news it's not Kingdom Of Deception LoRa.
KOD data set is huge (600+ image), so i decided to make LoRA's with smaller data set, because it's cheaper and less time consuming. So i can iterate this LoRA faster and understand how improve KOD LoRA.
In case you interested:
Game: https://f95zone.to/threads/fallen-makina-and-the-city-of-ruins-another-story-kagura-games.9662/
LoRA:
Bad news it's not Kingdom Of Deception LoRa.
KOD data set is huge (600+ image), so i decided to make LoRA's with smaller data set, because it's cheaper and less time consuming. So i can iterate this LoRA faster and understand how improve KOD LoRA.
In case you interested:
Game: https://f95zone.to/threads/fallen-makina-and-the-city-of-ruins-another-story-kagura-games.9662/
LoRA:
You must be registered to see the links
You don't have permission to view the spoiler content.
Log in or register now.
I don't know how to fix this for 100%) But in some cases you can get generations more stable, for example for Sabia+Neve scenes, if Sabia appears with pointy ears, then delete all pointy ears tags from positive, and add "pointy ears" in negative. It doesn't hurt so much, because Neve has in description tag "elf", which implies pointy ears. This should increase probability of getting Sabia with normal ears and Neve with pointy ears.
Also this works for scenes with Sabia+Orc or Sabia+Orcs.
Below two images generated with new model assistance, for both i tryed to recreate game scenes. As a result, win for first scene and fail for second (at least i something interesting in exchange).
You don't have permission to view the spoiler content.
Log in or register now.
I suppose that for specific scenes with more then 2 characters we need to train concept LoRA's.
- May 18, 2020
- 112
- 226
I used automatic1111 with forgeUI with Adetailer enabled for hands, body, and face. I also used the resolution of some of the actual KoD images. I was using automatic1111 a year ago before I deleted it and this brought me back lol.
After a while, i managed to put character from "Fallen" model, into scenes from "KOD" model, and vice versa, with decent quality.
Recommended strength:
first case "Fallen"=0.3, "KOD"=0.7,
second case "Fallen"=0.9, "KOD"=0.3.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
Recommended strength:
first case "Fallen"=0.3, "KOD"=0.7,
second case "Fallen"=0.9, "KOD"=0.3.
- Jul 12, 2017
- 1,154
- 1,104
Stable Diffusion 3 is out:
Look like this model is censored and also have licencing pitfalls(
You must be registered to see the links
Look like this model is censored and also have licencing pitfalls(
Last edited:
Well, i was tested ControlNet with screenshots from porn, and it worth it, but still you get approximate result and lot of garbage) I hope we get better refence_only model someday. Style is not preserved, and quality depends on screenshot resolution. Also it's nearly impossible to reproduce exact pose, but with this method possible to get close views like in image 23.png, which is impossible with txt2img method.
Other methods: canny, depth, open pose is bad, if they gives garbage as initial image, then you probably can't change it with prompt, at least in my case.
Making these images took me 5 hours of time. After all this time i think that Stable Diffusion shoud be renamed in Ustable Diffusion)
Oh, by the way, ADetailer is also dissapointing, this extension likes to generate breasts instead of faces)
Other methods: canny, depth, open pose is bad, if they gives garbage as initial image, then you probably can't change it with prompt, at least in my case.
Making these images took me 5 hours of time. After all this time i think that Stable Diffusion shoud be renamed in Ustable Diffusion)
You don't have permission to view the spoiler content.
Log in or register now.
Oh, by the way, ADetailer is also dissapointing, this extension likes to generate breasts instead of faces)
- Jun 9, 2019
- 344
- 1,335
When it comes to controlnet, depth or canny are cleary superior when it comes to reproduce a picture and transfert style. I suggest you lower the contol weight to 0.7, and keep it on balanced. Each pictures will be different as each seed are not made equal. But it more or less gives you what you want.
For Adetailer, it nevers gave me face for boob and what not. I suggest you go on the adetailer option, detection, and put the slider threshold at 0.7 at least. That way, adetailer won't detect faces on wall for exemple.
I use the Face_Yolo8s.pt for face, with a threshold of 0.7 and mediapipe_face_mesh_eyes_only for eyes, at 0.7 as well. Never had an issue.
Describing as much as possible the image in controlnet instead of what you want help too.
Exemple of controlnet with depth at 0.7 balanced :
You don't have permission to view the spoiler content.
Log in or register now.
And SD3 sucks apparently, not only it is censored, it also overhyped.
- Jun 28, 2017
- 312
- 504
Honestly, 5 hours to produce so many high quality images is really good production wise despite wasting a lot of time to filter out the garbage. With Mamaragan suggestions I think the workflow can be improved even more. This is quite amazing actually as it would allow to tell a story/scene based on porn comic books or movies and incorporate it into KoD.
Guys, guys! This is important!
When a woman tells you that "it would be fun to wake up to oral sex", this is NOT what she meant.
I'm confused
. As depth model in controlnet you using "controlnetxlCNXL_tencentarcDepthMidas", but this model for T2I_Adapter module and not for Depth module! Maybe i just didn't understand something... I tyed this model in Depth module, and it works, so i tryed some other (designed for other modules) and they work too!)You don't have permission to view the spoiler content.
Log in or register now.
Also depth_midas preprocessor produces different depth mask in my case, but not depends on model.
Should create another spoiler in firts post with links for extensions, preprocessors, diffusers and so on?
UPD 1: Added example image and setting to be more clear.
UPD 2: Actually there is high posibillity that "diffusers_xl_canny_full" and "thibaud_xl_openpose" doen't work with depth module or they contribution is small. I personally recommend to use "diffusers_xl_depth_SIZE", where size is: full, mid, small; depending on your amount of VRAM.
Last edited: