Guide Tools Fake Noob's Guide: NSFW Local Image Generation of a Character With a Consistent Face/Body, and Creating NSFW Image-to-Videos With Those Images

[Crazysteve122]

Decided that I should post my guide I made for another thread on here. The end goal will be videos of our chosen subject from our locally generated images along with some other tools I recommend. This is meant for noobs and is not a very technical guide.

- IMPORTANT: If you already use ComfyUI or other programs, part 3 of this guide and the dzine guide linked to in part 3 are great adjuncts to your repertoire of tools!


A Guide to Stable Diffusion Local Image Generation with ForgeUI PART 1

Preface
- The end goal of this guide is to get a consistent character using images of our choosing. It will be multiple parts.
- This part is to get you set up with local image generation, the consistent face/body will come later.
- You have a choice with local image generation. ForgeUI or ComfyUI. ForgeUI is simple, runs on older hardware, and is much quicker to get into than ComfyUI. If you're interested in ComfyUI then I can't help you. If you go around searching for UI's, do not use A1111's UI; It's outdated and ForgeUI is built upon it.
- This is a dummy's guide. I will assume you have 0 knowledge pertaining to all this. I also won't be going into too much technical detail and won't cover every single thing it can do.

Min Specs: If you have a laptop or PC that does not have a graphics card, forget it. Ideally you should have:

- 8gb vram Nvidia card with cuda, so basically anything modern within the past few years. You can run it on a 4gb vram card but the program will start digging into your CPU and you will experience slowdowns in your pc/generations while generating images; it's still completely viable.
- An SSD drive to install it on. Externals work just fine. Have at least 100gb free, more allows for you to install more models.
- Any good modern CPU
- 16gb of ram

Installation
- You first need a version of ForgeUI to install. I recommend the one below if you are illiterate with coding languages in general, like python. You don't have to mess around with cmd or all the other technical stuff as much. If you choose to install another version then look at the documentation for installation and use ChatGPT if you get stuck. I can't help you with installing other versions, but they all work relatively the same once installed. This version in particular is not being updated anymore, so if that bothers you search for a version that is still maintained.
- Head on over to https://github.com/lllyasviel/stable-diffusion-webui-forge
What to look for

- Download the recommended one and when you unzip in your desired install location, you'll see a folder with files.
- Run update, then run. If nothing launches or you get an error after using run, launch run again and it should open up the webui with your browser of choice.

Getting Situated
- When it opens up you'll be greeted by this. You'll also have a cmd window running, don't close it. What you'll see:

- Calm down and we'll go through everything you need to know, just the basics.
- At the top left click xl, let's go over why real quick, and install a model before we move further.

Base Model
- Stable Diffusion has multiple models to use, let's talk about the big ones.
- SD1.5: The OG. Much older but it has lots of models. Pros: Use models based on this if your PC is shit or you just want to generate stuff extremely fast. Cons: 512x512 based, can do larger images but too large and you get degradation without proper upscaling. Poor prompt adherence.
- SDXL: Souped up SD1.5. Pros: Good realism, 1024x1024 based and can do smaller or larger. Cons: Slow if your PC is shit.
- Pony: Anime heavy versions of SD1.5 or SDXL
- Flux: The newest model renowned for realism and details. Pros: Unmatched detail, hailed as the best. Cons: Slow, resource heavy. Has a tendency to produce plastic looking skin.
- We will be using SDXL because it's the best of both worlds. It's wedged in-between 1.5 and Flux in terms of speed and details.

Finding Models
- Go to Civitai, make an account, and make sure you allow nsfw models in your profile options.
- We will be using https://civitai.com/models/133005/juggernaut-xl
- It is based on SDXL
- Download it and drop it in (YourDriveName):\webui_forge_cu121_torch231\webui_forge_cu121_torch231\webui\models\Stable-diffusion
- Your path may be different, but the last 4 locations will be the same.
- Feel free to try any other models, I have tons myself and they all excel at different aspects. Those explicit AI porn videos I made in earlier posts were fed images made by pornmaster. If you want sexual acts use a sex based model.
- Do download pornmaster though as we will be using it in a part 3 later.

Your First Generation
- Go to settings and reload UI, or quit out of the browser and cmd then hit run again to get the model to show up. It will be under checkpoints.
What it should look like:

- Your GPU weight should only be lowered if you get an error in your cmd saying you're low on space.
- Prompt: What you want to see
- Negative: What you don't want to see
- DPM++ 2M Karras is pretty standard and works well, but you can play around with them all.
- 30 steps will get us a good image. Some models require more or less.
- Width and Height is pixels. If you plan on feeding an image directly to a video generator make sure it lines up with a commonly used ratio, like 3:4 or 9:16.
- Batch count vs Batch size: Size is images generated at once. Set to 1 if your PC sucks. So if you set batch count to 4 and size to 1 it generates 4 images in a row. Batch count to 4 and size to 4, you get 4 sets of 4 images generated in a row.
- CFG: How closely the AI follows your prompt. Lower = more creativity. 5-7 is generally recommended for most models. I run 7 most of the time. Play around with it.
- Test this out
Prompt: A Korean woman, wearing a black crop sweater, high heels, knee high stockings, black thong underwear, on an empty street sidewalk at night
Negative: BBW, fat woman, worst quality, low quality , lowres, illustration, 3d, 2d, cartoons, anime, painting, CGI, 3D render, bad anatomy, sketch, photoshop, airbrushed skin, overexposed, watermark, text, logo, label, blurry, depth of field, bokeh, blurry foreground, ugly nails, ugly fingers, fused fingers, missing finger, extra fingers, old, big breasts, bad teeth, shiny eyes, bad eyes
- I got this:

- You didn't because the seed is random every time when set to -1. Infinite possibilities.
- Click the folder icon below the image to access the image folder where they all save.

- PROTIP: Look at the cmd window to see the behind the scenes as it generates. Be on the look out for errors as well.

Other Tabs
- img2img: For editing already existing images. You can edit the entire image at a time, or just one part of it using inpainting. This tutorial would be insanely long if I went into all of this. Check out YouTube.
- PNG Info: Drop a created image in here to see all its metadata. Useful if you want to replicate the prompt and negative. Reusing a seed is also powerful for seeing how changes in steps or other options affect an image. Copy paste or hit a button to send it your tab of choice.
- Settings: Self explanatory
- Extensions: Self explanatory, some might be broken though if the GitHub page is down. Installing a broken extension can brick your ForgeUI and force a reinstall.

In Closing
- In part 2 I'll go over installing and using the adetailer extension for better faces, as well as creating what is called a lora to get a consistent face and body of a character. Part 3 will dig into video generation.
- Just this guide alone will allow you to go nuts with generating images however.

 

[Crazysteve122]

A Guide to Stable Diffusion Local Image Generation with ForgeUI PART 2

Adetailer Extension
- Now that we have ForgeUI and some models set up, let's get to work setting up adetailer.
- Adetailer is an extension that automatically scans the image for a face and upscales it.
- Go on over to the extensions tab.
- Click the available subtab. If it fails to load keep trying until it all loads up.
- Search for adetailer and install, it may take some time.
- Now click the installed subtab, and hit apply and restart UI.
- When you go back to txt2img you should a tab labelled adetailer right under the seed area. Click the checkmark to enable it and open it up.

- Note how it has it's own prompt and negative. I've typed in detailed portrait to tell adetailer to focus on the face, along with some stuff for the negative
- I use the 2nd model because it's supposedly better, but the first is faster.
- One more thing to change, open up the inpainting tab, go to "inpaint only masked padding, pixels" and set that higher, I use 128. The inpaint only masked should be checked by default, if it's not check it. This let's adetailer see more of the surrounding when working on the face and results in better generations.


- PROTIP: Adetailer supports hand inpainting as well, and some other features. There are guides online that go over each model. Use the 2nd, 3rd, and 4th tabs to apply other models to your generation.

Rerunning Our First Generation
- Find the image you used in the generation the first time, should be located at:
(YourDriveName):\webui_forge_cu121_torch231\webui_forge_cu121_torch231\webui\outputs\txt2img-images
- Drop it into PNG Info tab and click send to txt2img
- Now, make sure adetailer is on and all the settings we went over and prompts are ready to go. Click generate.
- Let's compare the faces.

- Way better

What is a Lora?
- In the simplest terms, a lora is like a miniature model that can insert a new idea into an existing model. This idea can be a person, a style, etc. We will be using it to generate a consistent character's face and body on ForgeUI.

Lora Creation Options
- Here's where things start to get muddier.
- If you want to create a free lora and lack the specs to do so, you'll have to scour google for a google collab that hosts a lora creator. There's plenty of links to such collabs on reddit, just spend some time searching via google. I've done it in the past and it came out okay. The collabs come and go and can be technical, so expect this method to take some time finding one and learning how to use it.
- If you have the specs and want a free lora, you'll have to look elsewhere for a guide to do so. My PC won't do it so I have 0 experience here.
- The easiest method will cost you $ via a site called Civitai.

Lora Creation Via Civitai
- Due to asinine laws the currency on Civitai called buzz, is split into green buzz, which can be bought like normal, and yellow buzz, which you have to buy via an instantly redeemable gift card.
- From what I know, both types can make a lora on there, it's just that only yellow can generate nsfw content on there. I had some leftover from my memberships earlier this year, so I used yellow, but green should work fine. Buy some buzz for $5 to get 5k, or buy a $10 membership for 10k, up to you. Note, loras cost 500 buzz to make, so you're covered either way.

Prepping your images, Full Body Lora
- Before we train we must gather our images.
- https://www.presize.io/ is a tool I found via reddit that lets us do this easily.

- Set the height and width to 1024.
- Now gather 30-40 images of your desired character and drop them in. Avoid pictures that are way too far away, have obscured faces (bangs are fine), and are of low quality.
- You can do it with less but it will not be a good lora. Do not do too many images as well.
- Crop them as necessary. Full Body does not mean cut their face off. Have their face in every image.
- You'll want a good mix of somewhat far, close up, face close ups, side profiles, different expressions, etc.
- 40 - 50% Front facing
- 30 - 40% 3/4 angle
- 10% profile (side view)
- These %'s are not set in stone. You can deviate somewhat and still be fine.
- Good lighting
- Varied outfits and settings. If you don't have naked or underwear pictures of the subject, the generator may keep putting clothes on them unless told to make them nude in the prompt. I do 1/3 nude, 1/3 lingerie, 1/3 clothed. This way it learns their body and doesn't guess nude shots.
- If you want a consistent body make sure to include plenty of body shots (hips/thighs to head) but don't overdo it, you can leave their legs out if need be, just make sure the face is clear.
- Once you are done download the zip.

Face Lora
- Open up a new link to that last site and drop in 30-40 images of them and crop to get them shoulder up. Remember, 1024x1024. The rules for front, 3/4 angle, and profile still apply.
- Download the zip.

Civitai Lora Training
- Go back to Civitai and in the top right hit your profile letter and on the dropdown, training.
- Click on published or private and you should see an option to train a lora.
- Hit image, character, and give it a name, next.
- Now drop your body or face zip in to make the first lora.

- IMPORTANT, your trigger word will be used to call the lora in ForgeUI. Do not forget it. Give it a name that is not recognizable. You don't want to name it girl, then your lora will mix up your character with all the training data a model used for a girl and it won't work at all. Use a random gibberish word and avoid numbers.
- PROTIP: Consider using the same name for your lora and trigger word so you never forget it.
- Click Tag, then autolabel.
- I set my tags to 12, but 10 is fine.
- Submit, and wait for it to tag your images.
- I could go into all the detail on what tagging means, but it's basically finding stuff in the picture to not associate with your character. If your person has brown hair in every image and you didn't tag brown hair, you'd have a hard time giving them blonde hair when generating images since it associates that with them. Your lora will still fall back to their base appearance just fine if need be, but tagging gives us flexibility.
- Hit next, then click on SDXL standard. Do not train it on a custom model. Training on the base SDXL model gives us the power to use our lora with any SDXL lora.

- Now train and watch the results flow in. You can set sample prompts if you'd like, but don't touch anything else.
- Sometimes the results look like unholy deformations. Don't worry and train your other lora (face or body).
- Once they're both done download the last epochs and drop them into:
(YourDriveName):\webui_forge_cu121_torch231\webui_forge_cu121_torch231\webui\models\Lora
- Boot up ForgeUI

Using your Body Lora
- If you'd prefer, you can follow this part and the next part in tandem. Doing so can make troubleshooting difficult, but can get you results faster.
- The hard work is done but the troubleshooting now begins. Your goal will be to find a suitable lora and trigger word strength.
- This is different for every lora, and can even be different for every model you use your lora with.
- Pick an SDXL model first under checkpoints.
- Hit the lora tab, and click on your lora.
- This will pop up in the prompt: <lora:YourLoraName:1>
- Now type in your trigger word
- So you get <lora:YourLoraName:1> yourtriggerword, the rest of the prompt

- Set your steps, cfg, etc. Leave adetailer off for now.
- Generate and see how it turned out. Shit? try changing that 1 by lowering it, 0.9 0.8 etc., or raising it, 1.1 1.2 etc.
- You can also change the triggerword strength using yourtriggerword(1.1), yourtriggerword(1.2), etc.
- The number in the <> affects lora strength, the () number affects how much attention the generator pays to your character, not strength just priority.
- This part takes a while but once you get a decent resemblance we will crank it up by x100 with adetailer.
- If your results suck: you trained your lora with bad images, not enough, too many, bad model use another, you used another lora creation method besides Civitai and you fucked it up, or your prompt sucks so keep it simple for now, make sure to specify front facing if you get a lot of profile or back shots.
- Examples of what you should get, my lora output as examples. Notice how the face is fucked still because it's not a close up image. Our face lora will fix it right up. You can technically use the full body lora with adetailer as well, and get a better face, but using a lora trained just on their face will look much better.
- If you're running this step in tandem with the next step, your results might look perfect already, or need a little tweaking.


The first image was made using JuggernautXL, while the 2nd was made with Pornmaster. (Since it's a side face shot, in adetailer we would add profile for an even better face for image #2)

Using your Face Lora with Adetailer
- Enable adetailer and use the mask settings we went over earlier.
- Since you can't directly insert the lora like in the main prompt, insert the face lora in the main prompt, copy paste the <> stuff to the adetailer prompt, and delete from main prompt.
- Add your trigger word and maybe say detailed portrait
So <stuff> trigger, detailed portrait in main prompt
blurry, distorted, asymmetrical, deformed, extra teeth, bad ears in negative

- Generate
- Shit? Well mess with the lora and trigger strength again!
- Experiment with changing the strength of the lora and trigger word in both the main prompt and in adetailer.
- PROTIP: Use the same seed and just change adetailer prompt settings. That way only the face changes and you can compare the same image side by side.
- Eventually you should have a banger of an image that resembles your character! You should be able to reuse these same strength values for that particular checkpoint model.
- If you don't: the seed you are using sucks so use another, you trained your lora with bad images, your prompt is ass, bad model, you used another lora creation method besides Civitai and you fucked it up

In Closing
- If you followed me this far congrats! You should now have a great lora for your character!
- Note that not all seeds will look like your character, some just suck.
- In your main prompt, try appending: masterpiece, best quality, highres, 8k, detailed skin, highly detailed to the end of the prompt for some improvements
- If you want a profile (side view) or ass shot with them turning to look back, specify side view, profile or body facing away with head turned back to look at camera in your main prompt, and be sure to add profile in your adetailer prompt! That way adetailer knows to focus on those 3/4 and profile shots from your lora.
- Part 3 will cover video generation.

 

[Crazysteve122]

IMPORTANT: As of 11/24/2025 Dzine's custom model has just released V2 and it is promising with NSFW content. Cocks go in and everything! Expect an update to this part of the guide once I have fully tested it! (Now updated)

Image to Video Generation
- By far the most popular nsfw image2video generator is WAN with loras. This requires a great PC and can be cumbersome and time consuming to get good results. 1 generation can take 20 minutes or more.
- You can access WAN with loras locally, you can refer to guides on here or elsewhere on the net for this.
- Civitai also supports WAN generation with loras, but it can cost you around 400-1000 buzz depending on your options, loras, WAN model choice, etc. It comes out to about $0.50 - $1.00 per generation, way too costly.
- We instead are going to focus on an AI platform called dzine. You can read all about it in my other guide.
- If you are not okay with spending money, then disregard the rest of the guide below. You're on your own for the video part, but at least you have consistent character images!
- IMPORTANT: DO NOT PURCHASE DZINE UNTIL YOU HAVE GONE OVER MY DZINE GUIDE LINKED ABOVE AND DECIDED, YES, I LIKE THIS.
- Dzine referral link for 50%+ credits. Your referrals helped me update this guide! I would have run out of credits otherwise.
https://www.dzine.ai/referrals/17524740339491ewixmNX

Image-to-Video Setup
- First, lets get back to pornmaster for explicit images.
- Using your new found knowledge from the past 2 guides, you should be able to get sexually explicit images of your character using pornmaster+your body lora and adetailer+your face lora. Remember to find the sweet spot for your lora <strength> and (triggerword) with each model.
- There are tons of models though, so see what you like most.
- PROTIP: Check out the images generated on Civitai and copy their prompts. Just make to sure to shove your own loras in at the beginning and change whatever you see fit (clothed vs. unclothed, location, anal vs pussy, etc.)
- You can even play around with the host of other tools Dzine has to create questionable SFW images, or use any other image really.
- Once you have your explicit images, or whatever other images you want to use, read the excerpt below from my other guide.

The following is also present in my dedicated Dzine guide.

Recommended Dzine Image-to-Video Models, NSFW and SFW
- There are 3 models in total that I recommend.

- Kling 2.1 STD or 2.5 STD (SFW): Will run you 60-70ish credits for a 10 second 720p video. With the 9k credits you get on a master subscription, you'll get around 120 videos, 180 for the first month if you use my referral code.
Overall: Best quality. Best consistency. Good speed (2-5 mins). Good for kissing, some fondling, dancing, butt cheeks if they aren't too exposed. Great for complex movements.

- Seedance Pro Light (NSFW): Will run you 14 credits for a 10 second 480p video. 30 credits for 720p. With 9k credits on a master plan you can get 640 videos at 480p, 300 at 720p. The extra credits from using my referral will get you 960 and 450 videos respectively for the first month.
Overall: Decent quality. Okay consistency as long as characters don't move too fast too quickly. Very fast speed (1-2 mins). Good for kissing, fondling, dancing, nudity, simulated sex, penetrative sex (although cocks don't go inside a woman very far), removal of clothes to show nudity is possible but can sometimes show underwear. Good for complex movements.

- Dzine Video V2 (NSFW): Dzine's custom model. Will run you 40 credits for a 10 second 720p video. With 9k credits on a master plan you get 225 videos, 337 with my referral code for the first month.
Overall: Good quality. Good consistency. Good speed (2-5 mins). Good for kissing, fondling, dancing, nudity, penetrative sex, blowjobs, removal of clothes to show nudity is possible but can sometimes show underwear. Okay for complex movements.
Note: Character's revert back to their original position for image2video. For this reason it may be difficult to get a lot of movement out of characters, so it works best if an explicit sex act is already in place to happen (cock near/in mouth or vagina). Do not prompt for a guy to walk in and fuck a woman, it doesn't work too well as the guy will walk back out and may be deformed. Works best if the male subject, whether it's a portion of him or his entire body, is already in place with the woman.

If I had to pick one model that's the best for purposes of gooning, it'd be Dzine's newly released V2 model.

Video Generation
- Let's go to the AI Video tab and check out all the models I talked about.

- Select a model, load your image, fill out a prompt, and check your settings.
- Reminder that you can use whatever images you create on Dzine, locally, or anything else.
- I won't be showing Kling examples since this guide is more for NSFW content generation.

Dzine Video V2 Examples

Using Dzine Video V2, 10 seconds, 780p
prompt: The woman gets fucked, the man's hands come out from bottom frame and fondle her butt

Using Dzine Video V2, 10 seconds, 780p
prompt: The woman fucks the black penis, the man stays still and doesn't come further into frame

Using Dzine Video V2, 10 seconds, 780p
prompt: The woman sucks the black cock, saliva drips from her mouth

Using Dzine Video V2, 10 seconds, 780p
prompt: The woman sucks the black cock, saliva drips from her mouth

Using Dzine Video V2, 10 seconds, 780p
prompt: The man fucks the woman with his big black cock, it slides deep into her pussy

Using Dzine Video V2, 10 seconds, 780p
prompt: The man fucks the woman with his cock, his hand stays still

More examples:
Reverse Cowgirl
Doggy
Cowgirl
Doggy 2
Handjob