Hello, fellow hubbers
So, it has come to my attention that some people would love to enter the field of AI Image generation, but don't know how to do that (one example I can think of is the last comment on image 198377, or this topic here
hypnohub.net/index.php?page=forum&s=view&id=106496).

While there are many tutorials on the Internet, I think it would be useful to discuss various programs, models, techniques, etc. specifically aimed at creating Hypno|MC-based content. I was thinking about writing a quick guide on how to set up a StableDiffusion environment and what kind of checkpoints\loras\prompts worked best for me so far to recreate these kinds of pics, so that all the other AIgeners can shi... share their own opinions on how they achieve what they create. Sharing is caring in the AI community, after all, <3... and I'm always up to learn new methods to improve and change my generations.

Would that be of any use to the fine people here? Is that allowed to be discussed?

I'd be very interested, thank you!

Guess we get nothing :(

I would be up for sharing my workflow.

But if you are just getting started and have a decent nVidia GPU I recommend installing Automatic1111 then downloading the AnythingV3 model checkpoint from civitai.com. Plenty of guides online for using it along with prompt suggestions and example image output from prompts on civitai as well. I think using SD1.5 is a good balance between compute requirements and output quality.

One trick worth sharing is that the output resolution and aspect ratio matters a lot more than you would expect. With AnythingV3 don't go over 768 in any dimension or you'll get distortion. But don't feel locked into 768 x 768 just because that is the "ideal" for AnythingV3 either. I like to start at 640 x 400, edit an output I like to clean up details, fix a pose, etc., and create a 1280 x 800 version of that using img2img. Results are much better because you are able to modify the composition of the 640 x 400 guide image before generating the real thing.

At this point, AnythingV3 is very out of date for a 1.5 model. If you're using 1.5, one of the Based models (65 or 67) is the way to go. Overall though, 1.5 models themselves are falling out of favour, being replaced by XL models like PonyXL and AutismMix. They're leagues better at composition and prompt comprehension and since they're trained on SDXL, you can gen at a much higher resolution (1024x1024).

That's not to say that XL is definitively better than XL though. For one, Controlnet models are much better on 1.5 than XL. CN allows you to have a lot more control over the initial generation so having an inferior version with XL means you'll be playing the slots a lot more. There's also the fact that there are *way* more loras trained for 1.5. The XL models are still fairly new so you may find it difficult to find a specific character or pose.

Have to agree with Guile there.
While 1.5 owned the title of being much more flexible than XL, this was recently shattered by the arrival of PonyXL, a model that (quite literally) gave out a new wave of AI generations. All my generations are made with Autismix DPO, and again, not joking, this checkpoint (alongside all other Poni Derivates) was a game changer due to how many tags were embedded in the models alone without needing LORAS. If you can muster enough horsepower with your GPU (I always suggest using ForgeUI over Automatic1111 for old ones), use these ones: they are much more "natural language" friendly, too, so you don't have to take the problematic process of learning prompt refining immediately.

However, as Guile mentioned, some poses are pure agony to get with XL and Pony compared to the old 1.5. Also, Pony and derivates are EXTREMELY picky with tags: just by adding a single word that it doesn't like SD can produce you a graphical abomination (Jackson Pollock style), and the same word could work perfectly fine with another prompt because... no one knows why. One example out of the blue: the term "fingering" for me is prone to disaster 99% of the times.

Also, as HypnosisIsGreat mentioned, resolution is the VIP in these things: you CAN technically create different outputs with different resolutions, but by following predetermined values (you can find them in here: www.reddit.com/r/StableDi...l_resolution_cheat_sheet/ . Print it and put it in your pocket) your quality will improve dramatically. If you also want to amp up the final product, consider using a Hires Fix (my way to go are Esgran Anime 6B for animeish stuff and 4x_NMKD-Siax for everything else; civitai.com/models/147641/nmkd-siax-cx) and, in case you use a Pony derivate, using a style LORA (like smooth_anime). Also add this extension to your UI (github.com/DominikDoom/a1111-sd-webui-tagcomplete): seriously, helps a lot.

As a final piece of advice to newcomers: go to Civitai, make yourself an account (which is needed to download checkpoints and the like) and start copying the prompts of us "veterans" and test them yourself. We all started to learn this way, and that's 100% fine. After that, try new LORA, new checkpoints, extensions like ADDetailer, ControlNet, and also new tools like Inpainting (and the recent addition of Soft Inpainting), and learn how to make masks on Digital Programs like Photoshop or Krita, go here (openmodeldb.info/) and get new hires models, seriously, there is SO much to learn but, to start, get a functioning UI and start generating your first abominations (yes, your first will not be pretty, but sometimes they will produce interesting results: for example, today I tried an action figure checkpoint, then switched back to an older prompt for Autis BUT forgot to switch checkpoints.... the result was still an interesting action figure in a decent pose and with decent background xD). Oh, right, and always remember to add clip_skip 2, to use VAEs properly and add "bad_anatomy" in your negative. And share, share, share your results alongside with the metadata when possible: seriously, sharing is caring in the AI community. We wouldn't be here at this point without all of us being open (most of the time) with our achievements and how we got there.

P.S.: If you are pondering which Graphical Cards to get, for the love of God, DON'T get an AMD one. I am an AMD veteran here with AI (7900XTX) who recently swapped to 2060 due to his GPU being fried by the PSU. The amount of headaches I got from trying to run programs on ArchLinux was infinite (AMD and AI do not work well on Windows yet) while, with this old Nvidia card, I just downloaded an executable, pythorch and *boom* works like a charm. Yes, it is true that the price difference is substantial (and I won't sugarcoat it... Nvidia is an ass), but if you have to choose and you have the money... consider Nvidia over AMD.

To add on, XL models are still somewhat newer. You can't use XL and 1.5 interchangeably, if you use a 1.5 checkpoint then all of your loras and the like need to be based on SD 1.5 too.

But yeah despite AMD cards having great VRAM, they are awful for a lot of AI applications. It's why NVIDIA is eating them alive.

Blame SOMEONE who decided to stay quiet (despite promising news) regarding the release of the needed ROCM for the 7000 series to run AI programs despite having the entire Github community wanting to help them as they did multiple times in the past. I bought my 7900XTX in February (2 months after the release), and I still had to wait 3 more to get Stablediffusion working (and only because one of my tech friends was able to bypass so many limitations: most people, even experts, were stuck).

Now, I won't deny that most of the programs in the world run with CUDA while ROCm is usually not so sure it will be implemented, but still...

Hey all, thank you so much for the info in this thread! I've gotten started with AutismMix, and I'm finding character's faces, specifically the eyes, to be very inconsistent. Sometimes they look good, but most of the time they look like eldritch abominations. Does anyone have any tips on getting consistent eyes and faces with AutismMix?

That's what I said regarding how extremely picky is the model xD Generally you should try to generate low steps ones first, with each one removing one word or another and see what changes: sometimes it's one word, sometimes it's one weight too much, sometimes it's even a word positioned wrong in the prompt.

If you want some help with the prompts, you can use some of mine in my profile (civitai.com/user/The_rule_of_Thetra) and try to recreate the basics for faces\eyes. One word of advice: Autismix kinda hates the term "hypnosis", probably because it is associated with spirals who, ultimately, fuck up the generation, so... don't use it.