[Stable Diffusion] Prompt Sharing and Learning Thread

[osanaiko]

To be hon

Thanks I have tried the regional prompter but it really doesnt work, I mean its one thing to have sun, moon, boy etc. but to partition it and have two people having sex turn out exactly like you want is not happening, LORAS get mixed, clothing is all over the place etc, after a lot of trial and error I decided to try my hand at in painting and that gave me so much better results with so less efforts just trying to work out certain chinks like blur, merging and proper way to mask when two characters are intimately involved. However If you do have any good guide for regional prompter for handling two people have sex better do guide me

To be honest, regional prompting is not a technique I've spend a lot of time on myself. I was parroting some advice I read elsewhere.

In truth, my process is to cut-n-paste the best parts of generated or real art images into a frankenstein's monster that is close to the composition I want, then use repeated cycles of img2img and manual editing +inpaint to finalize the result.

> "turn out exactly like you want is not happening"

This is precisely how the experience of using diffusion image gen goes...

 

[Sepheyer]

To be hon


To be honest, regional prompting is not a technique I've spend a lot of time on myself. I was parroting some advice I read elsewhere.

In truth, my process is to cut-n-paste the best parts of generated or real art images into a frankenstein's monster that is close to the composition I want, then use repeated cycles of img2img and manual editing +inpaint to finalize the result.

> "turn out exactly like you want is not happening"

This is precisely how the experience of using diffusion image gen goes...

I am starting to think the answer is to move away from SDxx to Wan VACE, hardware permitting. That's what comes with "built in" hand fixes, regions, characters, consistent characters, muuch stronger context recognition.

 

[JhonLui]

Just found this thread.

Take this from a "noob", so it may be more of a question than an answer..
but I'm facing the same problem with a different approach: (I'll try to be brief).

First a couple of considerations (maybe superfluous):
-WAI nsfw is an extremely good and complete base, so, before using a Lora, always check if the character is already in the Checkpoint model, than force the original style by adding "anime screencap" for anime or "official art" for 3d. Also the character must be specified with the trigger word WAI uses (there should be a database on the Github page) example:
"Nico Robin" is fine, but Nami has to specified with the series :"Nami \(one piece\)" or "Nami_(one piece)" to get a realistic look.
This yet doesn't solve the "clothes issue".
- For WaiNsfw (at least in HD resolution 1024x1024 etc..) 20 to 24 steps are more than enough, after that either you waist time, or there is a further risk of mismatch.
- Lately a couple of addons/alternatives came up you might wanna try:
"wai-2Rectified-V140": a lora that "focuses" Wai allowing a quality consistent result even with Hyper or Lightning (4steps)
"animeScreenshotMerge_v32": a Checkpoint based on WAI but "specialized" in multiple characters display (didn't try it a lot, but seems to be pretty good).
"koronemixVpred" might also be worth a try, as the VPred Models are still in developement, but seem to be much more precise.

Now onto the method (not a solution but a different approach that may give insights):

Since I use Hyper 4steps (with Fooocus) I tend to use the "strenght in numbers" approach: so start with a basic prompt, than when found the desired overall look, fix the seed or put a high Guidance (>7) and start building up the prompt untill the desired result.
Now.. in Foocus it's a mess specially with different chars because the structure isn't "language logic":
subject/description/action/object/description/reaction/situation but more "situation/description/subject/object/description/action/reaction/details, so switching characters or prevent a mismatch means a prompt overhaul...
But in Forge the "Break" method should still work, you may wanna look into it. (not sure)

As for something close to a solution to your problem.. I'm afraid the only way is to switch to ComfiUI (or SwarmUI: basically the same but with a Forge interface for start) the only way to have a detailed control on separates Loras ..afaik.
[both + many others free, portable and offline capable with Stability Matrix, provided you have a 1TB SD drive to dedicate]
Than use Control.net models (CPDS) for direct posing or inpainting. Yet I can't give you specific directions because I like to "play casually" with AI.

I hope it can be of at least some help, and any further detail or explanation is most welcome to help me too.

 

[mosnew]

I am starting to think the answer is to move away from SDxx to Wan VACE, hardware permitting. That's what comes with "built in" hand fixes, regions, characters, consistent characters, muuch stronger context recognition.

first time i am hearing of this

 

[mosnew]

Take this from a "noob", so it may be more of a question than an answer..
but I'm facing the same problem with a different approach: (I'll try to be brief).

First a couple of considerations (maybe superfluous):
-WAI nsfw is an extremely good and complete base, so, before using a Lora, always check if the character is already in the Checkpoint model, than force the original style by adding "anime screencap" for anime or "official art" for 3d. Also the character must be specified with the trigger word WAI uses (there should be a database on the Github page) example:
"Nico Robin" is fine, but Nami has to specified with the series :"Nami \(one piece\)" or "Nami_(one piece)" to get a realistic look.
This yet doesn't solve the "clothes issue".
- For WaiNsfw (at least in HD resolution 1024x1024 etc..) 20 to 24 steps are more than enough, after that either you waist time, or there is a further risk of mismatch.
- Lately a couple of addons/alternatives came up you might wanna try:
"wai-2Rectified-V140": a lora that "focuses" Wai allowing a quality consistent result even with Hyper or Lightning (4steps)
"animeScreenshotMerge_v32": a Checkpoint based on WAI but "specialized" in multiple characters display (didn't try it a lot, but seems to be pretty good).
"koronemixVpred" might also be worth a try, as the VPred Models are still in developement, but seem to be much more precise.

Now onto the method (not a solution but a different approach that may give insights):

Since I use Hyper 4steps (with Fooocus) I tend to use the "strenght in numbers" approach: so start with a basic prompt, than when found the desired overall look, fix the seed or put a high Guidance (>7) and start building up the prompt untill the desired result.
Now.. in Foocus it's a mess specially with different chars because the structure isn't "language logic":
subject/description/action/object/description/reaction/situation but more "situation/description/subject/object/description/action/reaction/details, so switching characters or prevent a mismatch means a prompt overhaul...
But in Forge the "Break" method should still work, you may wanna look into it. (not sure)

As for something close to a solution to your problem.. I'm afraid the only way is to switch to ComfiUI (or SwarmUI: basically the same but with a Forge interface for start) the only way to have a detailed control on separates Loras ..afaik.
[both + many others free, portable and offline capable with Stability Matrix, provided you have a 1TB SD drive to dedicate]
Than use Control.net models (CPDS) for direct posing or inpainting. Yet I can't give you specific directions because I like to "play casually" with AI.

I hope it can be of at least some help, and any further detail or explanation is most welcome to help me too.

will try that as alternate...i think my original point got lost. i first make image i like as you suggested with minimal prompt, use the chance to define my first main character and focus.. ie female and then change the second character, I had very decent success with inpainting. what I am trying is to sort out chinks that cant be found on youtube video as its SFW like best way to select a mask, ensure consistent changes like background, action, certain parts of clothes from where the character is swapped out etc.

 

[JhonLui]

will try that as alternate...i think my original point got lost. i first make image i like as you suggested with minimal prompt, use the chance to define my first main character and focus.. ie female and then change the second character, I had very decent success with inpainting. what I am trying is to sort out chinks that cant be found on youtube video as its SFW like best way to select a mask, ensure consistent changes like background, action, certain parts of clothes from where the character is swapped out etc.

I kind of guessed it, but couldn't take it for granted...
In that case (afaik) the impaint method seems to be the only way, with Control.net for adjusting the character, but at this point I wouldn't consider WaiNsfw + EulerA the best option since they are the most "powerful" but tend to get messy.
I suggest you try the Checkpoints I mentioned for anime, StableYogi's or Cyberillustrius for realistic, and 5MoonDoll for 3d.
Also BeretMixReal is excellent for realistic or 3d for "eastern look" characters.