Getting Started with AI Gen

AI image and video generation has come a long way. The bar for entry has never been lower, and you, yes you, can with a little bit of effort make your dreams come to life. I wanted to share a summary of some of the things I've learned starting out on this journey and encourage all of you to give it a try if it is something you are interested in.

• AI gen is very random and expect it to work like a partnership between you and the AI. You are more of a "producer" trying to help guide an "artist" (the AI). So, you may find it challenging to get exactly what you want either the first time around, or ever (without the use of training your own models). That said, if you can accept the randomness and let the AI surprise you a bit, you can probably generate some good stuff with minimal effort. But, be aware, anything you add to the prompt can mess with the randomness. For example, adding a steaming teacup to my Mom Balls images resulted in a change of her hairstyle and makeup. Doing too complex of prompts can create major issues leading to deformation and distortions that'll degrade into horror images. Messing with the CFG and Steps can also give different results. It is best to experiment iteratively and gradually. I highly recommend using an LLM chatbot to help refine your prompts. I use x.com's Grok and Opera's Aria myself, but ChatGPT or DeepSeek should work too. Just put something in like 'Help me refine this prompt for Stable Diffusion: "blah" with negative prompt: "blah"'. Sometimes it'll cut out too much, but can really help you clean up a prompt that has gotten too messy or complex.

• Using models: Checkpoints and LoRAs are key to generating quality NSFW images. See tools below. CivitAI is great because you can see what others are using to generate their images, including their prompts, LoRAs and Checkpoint. For now, all you need to know is that a Checkpoint is sort of a base model and is the most important piece for generating your images. Understanding how to use the checkpoint is key and so if you aren't getting the results you want, I'd look up the checkpoint you are using and consult a prompt construction guide. Same with LoRAs, these are smaller models that run on top of the base model and can alter the way it generates images based on key words. Make sure you look up the LoRA to see what words trigger it. You may want to use a LoRA for example to generate images of huge futa dicks in a certain style. Make sure you are using the right word in your prompt so the LoRA will trigger. You can mess around with the LoRA's strength in order to make the final images look more like the images the LoRA was trained on.

• Starting out, I'd recommend using free Cloud-based SAAS tools to do your AI gen. There are a lot of great options out there, here are a few that give you free credits:

  • https://civitai.com - this is a fantastic place where you can see what others are working on, see how they generated the work, and even "remix" them. This is how I got started, finding images I liked and tweaking their prompts to be my own. You'll probably run low on their currency, Buzz, before too long and may want to get a subscription. But, you can get a fair bit of free gen doing basic image and video generation as long as you are posting, following, and liking images/videos. Recommend doing each of these activities at least once a day to maximize return, plus claiming your free Buzz in the generator. I still post from time to time here to give back to the community (you can remix works I've posted here to your liking with a few clicks): https://civitai.com/user/momballs . Another big benefit of CivitAI is that they provide downloads for models, checkpoints, LoRAs, all that stuff you'd need to switch to local generation. But, you can also try all these things out in their generator which I still use from time to time since it is sometimes faster than downloading these big files and getting them into my ComfyUI workflow.

  • https://www.klingai.com - Kling is probably the best video generator out there right now. Its advanced physics and large training model give its videos second-to-none realism and prompt-following capabilities. Problem is: it is now heavily censored. But, a free subscription will give you daily credits for login. But, you must use them that day or they will expire. I generate a fair amount of "softcore" stuff in Kling and probably will continue to even after I get Hunyuan working for XXX stuff. CivitAI can also generate Kling videos using their API, spending Buzz rather than Kling credits, but if you are going to generate Kling video with CivitAI, upload your image for image2video to Kling first to see if the censors reject it, it'll save you some time. Speaking of which, that is my process, I generate images with Stable Diffusion (now using ComfyUI, but previously using CivitAI), then use image2video techniques to bring those images to life in video. Getting past censorship is difficult in Kling, most images with even partially exposed genitals or nipples will get flagged. But, sometimes you can get lucky. But, genital bulges and large covered breasts can usually get through. Because it isn't trained with NSFW images, it may have difficulty handling the physics related here and you may get very inconsistent results. I have a big "outtakes" reel I'm working on with some of the silly results I've gotten.

  • https://app.pixverse.ai - PixVerse is a bit like Kling, but a lot less likely to listen to your prompt. But, you can probably generate a couple videos a day for free with it. As above, you'll run into challenges with censorship. And its very liberal interpretation of your prompt. My recommendation is to keep the prompt extremely simple like "woman dancing" or "woman drinking tea" so it doesn't create something crazy surreal (though often hilarious).

  • https://www.vidu.com - Vidu is like a budget Kling. Unfortunately they don't provide much in the way of free credits (and no daily free credits), but it is fairly cheap to get a subscription. Censorship is an issue here, but I've found it is a little more forgiving than Kling or PixVerse's, allow you some testicle and penis cleavage.

• Moving to local generation and beyond

  • The above tools should be enough to get you started for cheap. Eventually, you may wish to try out more advanced tools, especially if you have the hardware. This'll also mean you won't have to pay AI service providers for subscription costs and can basically generate to your heart's content for free. Local generation requires a fairly beefy graphics adapter with plenty of VRAM. My RTX 4060 with 8 GB is fine for generating images but can't create quality video.

  • The premier tool for local gen is ComfyUI. Note, this is a very complex tool, but it does let you have complete control over the generation process and is highly extensible with many folks out their creating custom nodes, scripts, workflows and more for it.

  • AUTOMATIC1111 or A1111 I've heard is like a beginners ComfyUI. It doesn't give you as much flexibility, but is much simpler. I haven't tried it though since I moved straight to ComfyUI.

  • For true NSFW video, Hunyuan is the most promising tool at the moment. The NSFW AI community is feverishly training LoRAs right now for this tool and it won't be long I feel before we'll all be generating excellent XXX video. But, the quality of your video generation is dependent on your hardware. Luckily, you can rent powerful GPU processing in the Cloud using tools like runpod.io and vast.ai. If you want to see what Hunyuan can do right now, check out https://x.com/Futasnacks .

• If you are interested in seeing how I've generated images and videos including the original prompts, Teacher's Balls subscribers can access my notebook (including some tips on post-production, tools, my task list, and even my super-secret ideas brainstorm which may hint at things to come) and project archive here, complete with some additional rejected content: https://www.patreon.com/posts/teacher-balls-122523045

• Lastly, AI gen has gotten where it is today because of an open source community. This is why I don't mind sharing my prompts and processes and tips and I encourage all of you to do the same and give back more than you take. We stand on each others' shoulders to make this all happen. Not only is there little use in hording secrets it is arguably counterproductive since information you share with the community can lead to new developments that you can then utilize freely.