- Oct 20, 2021
- 304
- 678
- 136
- Jul 4, 2017
- 3,350
- 6,436
- 707
I'm new to AI image generation as in I literally started today (though I do have a lifetime of experience tagging booru boards if that helps) and I just installed Pinokio (someone recommended it as an easy way to start learning to generate stuff locally). Through it I installed Foocus to fiddle and learn with image generation and Wan 2.1 for video generation (tho I'm definitely way more interested in image generation right now). I also installed kohya_ss (a Stable Diffusion LoRa & Dreambooth WebUI) cause I really want to generate anime and/or western hand-drawn style sex scenes with consistent characters, and from what I gather I need to learn the LoRa stuff to do that. I haven't fiddled with it yet cause I kinda don't know where to start so I think I need to watch tutorials or something.
Though I mostly want to use already established characters for those scenes (like, say, Raven from DC, Hatsune Miku, D.Va from Overwatch, Princess Peach or whatever), so I probably need to learn where to download already trained datasets or something and where to input them. I'm also much more interested in generating actual sex scenes than sexy posing images (nothing super complicated in terms of number of participants or acts: hetero sex, fellatio, occasionally anal, etc, but with some very specific details like e.g. I don't want huge breasts but I do want huge dicks and the dude to be massive, occasionally some ahegao face, stuff like that).
Anyway, before I dive headfirst into learning these tools and setups, is there something you recommend me to change? Maybe Pinokio is not recommended at all or you think I should swap some tool for another or that I should go to X specific tutorial or whatever. I do only have a 3050 with 8GB VRAM and 16GB RAM.
Thanks in advance to anyone that has some tips!
Though I mostly want to use already established characters for those scenes (like, say, Raven from DC, Hatsune Miku, D.Va from Overwatch, Princess Peach or whatever), so I probably need to learn where to download already trained datasets or something and where to input them. I'm also much more interested in generating actual sex scenes than sexy posing images (nothing super complicated in terms of number of participants or acts: hetero sex, fellatio, occasionally anal, etc, but with some very specific details like e.g. I don't want huge breasts but I do want huge dicks and the dude to be massive, occasionally some ahegao face, stuff like that).
Anyway, before I dive headfirst into learning these tools and setups, is there something you recommend me to change? Maybe Pinokio is not recommended at all or you think I should swap some tool for another or that I should go to X specific tutorial or whatever. I do only have a 3050 with 8GB VRAM and 16GB RAM.
Thanks in advance to anyone that has some tips!
art4vandelay
Member
- Feb 2, 2025
- 148
- 963
- 152
JhonLui
Well-Known Member
- Jan 13, 2020
- 1,146
- 1,136
- 284
That is your limit, as AI generation uses the V-ram, so untill you can upgrade you will have to start small.
- Also.. overclock the V-ram can give you a 5% speed boost -
Whatever you choose to use as a program you are stuck with Fp8 models (checkpoints) and SD1.5, SDxl, Pony or Illustrious Loras, unless you want risk to take 3 minutes to generate an image just to find out the girl has 3 legs... and forget the videos completely. Not because you can't, but to avoid hours lost, discomfort and frustration.
If I have to suggest... well I use Stability Matrix as platform (automatic installation of required components when needed, portable and offline, + shared models between programs and internal Civitai downloader).
As Inference: Fooocus Mashbit for quickies, or SwarmUI (ComfyUI with Fooocus interface) for Flux models or other complex stuff like handling 2-3 different characters precisely.
The only downside is that you have to dedicate (strongly suggested) a 1TB m2 driver just for it, as the collection tends to grow fast. [I'm at 1400+ models for 650GB..].
For your specific needs, I'd use a good versatile model (checkpoint) like WaiNSFW that can read both Illustrious and Pony Loras,
in coop with WAI-2Rectified Lora (a miracle refiner), or the latest Stable Yogi for Realistic. Than just search for your preferred style Lora or a specific character, if they are not already included in WaiNsfw.. there are tons.
Than you can choose to try newer models, there are very good ones lately all with specific attributes (Beret Mix, Korone Mix, Anime screenshot Merge, Luminar etc...).
Also look for Comics Loras, not to use them, but to learn how they use prompts
That should allow you to generate already good images in HD in 8-10 seconds in SD Hyper (or ~1min for 30 steps) while you gain some experience.
4 seconds Fooocus-SDHyper (4steps) - 4060Ti 8 GB - No effects- WaiNsfw, Euler A - Loras: Yashin Nami, Wai Rectifier (0.3) - No negative promptHope it helps.
Last edited:
Hello everyone here, first post since I've first discovered F95. I'm also starting my journey with Stable Diffusion, for various reasons (yeah, right ^^;
. I'm slowing upgrading my PC configuration. For now, a Nvidia RTX 3060 with 12 Go of VRAM would be a sweet spot, financially, quality and performance-wise. 32 Go of fast enough RAM would also prevent too slow generation. Yet I'm still new when it comes to find the right models. One modest goal would be generating generous images for a small VN that I'm writing as hobby. Specially with the scenes where they would be... action.
JhonLui
Well-Known Member
- Jan 13, 2020
- 1,146
- 1,136
- 284
Hi. I take your question as the occasion for making a small update on what I wrote above.
Regarding your needs, it's only about how much money you have to spend... and a couple of other things.
- Forget the RAM and focus on the Vram (GPU):
Ai generation is made to use only the graphic card, than it can "fall back"on Ram if the loaded models are exceeding the Vram, but than the generation will be fully performed at the slowest speed available and take ages... times 8 or 10.
(cpu: DDR4:2800mhz; DDR5 4400Mhz - Gpu: GDDR6 9000Mhz; GDDR7: 11000Mhz)
[I should use the iterations/second values (much bigger difference), but Mhz are more commonly understood].
So spare the money on the Ram and buy a better MoBo (pcie5+Nand M2) and/or Gpu.
- 12GB Vram is kind of.. in the middle?:
It's better more than less.. but the situation is different from the videogames where there is an actual problem (because the devs are deft... but hopefully will be solved by texture compression eventually).
The vast majority of dictionaries (Vae) which are always fully loaded is fp8 (7.3GB) or fp16 (14.2GB) than 32 and so on...
I might be wrong but I don't recall seeing models for 12GB, so you would be stuck with fp8 models anyways or risk to fall in the abovementioned case. (doublecheck on that, specially for GGUF models).
In this optics the Intel Ark 16GB could represent the best budget option for AI, but than the computer is ment to be an all-arounder machine, that's why (no matter what the youtubers are payed to yell) Intel LG1700 +Nvidia (+Shitdows) is still the way to go in my opinion (budget, compatibility, stability, productivity, multy-task performance bla bla bla..).
Don't take it for granted, since the situation is litterally changing every day, just be safe and compare the numbers, as many promoters/reviewers often "forget" to mention the weak spots. In any case, if you can, go for a 16gb card.
For who is still not asleep... now the small update:
There is a new version of the "Wai Rectifier" Lora (A14), which is lighter-effect but much more precise, and now works well also with SDXL Lightning 4steps (twice as fast than Hyper in Fooocus) with more than acceptable results (if you're using a good quality model/lora obviously).
Last edited:
- Jul 4, 2017
- 3,350
- 6,436
- 707
Vilkas91
PONY XL is a well regarded checkpoint (i.e. SDXL base model + custom training) that has good capabilities for making images with various sex positions.
Make an account on Civitai.com and try this search URL:
You can find the base Pony XL and also further finetune models that focus on PonyXL + Anime or + Realismn etc.
PONY XL is a well regarded checkpoint (i.e. SDXL base model + custom training) that has good capabilities for making images with various sex positions.
Make an account on Civitai.com and try this search URL:
You must be registered to see the links
You can find the base Pony XL and also further finetune models that focus on PonyXL + Anime or + Realismn etc.
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,703
- 4,184
- 448
My dudes, WAN 2.2 is the tits. Zip file contains the workflow. This is image-to-video, eight minutes to generate, needs a few large models downloaded (about 20GB total).
View attachment a_00018_.mp4
View attachment a_00007_.mp4
You don't have permission to view the spoiler content.
Log in or register now.
View attachment a_00007_.mp4
Last edited:
Sharinel
Active Member
- Dec 23, 2018
- 609
- 2,440
- 448
Shows as a png file, 'open in new tab' show as a png file, copy into comfy and shows as an avif image and has no workflow
[/QUOTE]
- Jul 4, 2017
- 3,350
- 6,436
- 707
[/QUOTE]
I think that's caused by the attachment server, which is auto transcoding images to avif for bandwidth savings.
Not sure when it started, but it obviously breaks the functionality of getting the metadata from a png txt block.
Doesn't work even if you leave a file as an attachment and don't include it in the post body. I wonder if it is due to the extension or if the server is doing some MIME type sniffing. If it's just the extension then we could add like ".txt" on the end of the filename to get around it.
Might be one for the F95 Adminn brains trust.
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,703
- 4,184
- 448
pu55y_sl4yer
Member
- Sep 11, 2020
- 109
- 384
- 183
Guys how do i use the img2img? When testing with realdosmic checkpoint, i just asumed that if i give it an hentai it would turn that into realistic but it looked absolutely nothing like the hentai at all. Help plz
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,703
- 4,184
- 448
You'd get a fast and useful answer if you reduced your ComfyUI workflow to the bare must-have bits and uploaded it here (in json/zip format) for others to troubleshoot.
Synalon
Member
- Jan 31, 2022
- 234
- 676
- 184
Put the image in, make sure the size matches, set the denoise (lower has less change, higher has more change, start at 0.5 and lower or raise to suit).
In the prompt sometimes you don't need much more than realistic or whatever type you want, other times you might need to describe the clothes.
Also the checkpoint may matter, a checkpoint that pretty much only does anime won't really work for realistic and vice versa.
JhonLui
Well-Known Member
- Jan 13, 2020
- 1,146
- 1,136
- 284
The f**k is "realdosmic" couldn't find it anywhere
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,703
- 4,184
- 448
QWEN is impressive AF.
It doesn't do NSFW out of the box.
Workflows - attached.
It doesn't do NSFW out of the box.
Workflows - attached.
|
|
|
|
|
|
Last edited:
Great images. The zip file contains 2 images - is the workflow embedded somehow?
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,703
- 4,184
- 448
Yes, the images are the workflow - open them in ComfyUI and it will be up (and for the missing models it will give you a dialog to download the models - as long as you have the latest CUI version).
Sepheyer
Well-Known Member
- Dec 21, 2020
- 1,703
- 4,184
- 448
Man, I struggled with SDXL - never really got into it, went back to SD.
So, I def recommend Qwen - pretty much right out of the box it works the way I expected SDXL to work
Fucking king.
Workflows attached.
So, I def recommend Qwen - pretty much right out of the box it works the way I expected SDXL to work
Fucking king.
Workflows attached.
|
|
|
|
Last edited: