|
Working with Raenir Salazar now. I only work with Stable Diffusion and image generation but follow all of this closely so I know of and can provide links to a large number of different tools. Sadly I don't play Trad Games so I don't know what will fit best here. It does sound like fun to play it's just going to be quite awhile until I have enough free time again. Could anyone give me some examples of the main games and ways people play? Sitting around a table with people, what's in front of them? Character portraits? Minis? Dungeon maps? All of these can be made with AI but if you give me a list I can better get specific links on how to do so.
|
#
¿
Jun 13, 2023 05:35
|
|
|
|
| # ¿ May 17, 2024 02:42 |
|
|
Kestral posted:Excellent write-up! Thank you for this, I have a much better understanding of things now! There are easy ways to start using some of these tools, like if you want to just dip a toe in the water you can make a quick character portrait, grid map, or scenery without much effort. There are all different styles of custom models and finding one that captures the gothic aesthetic for Vampire: The Masquerade or gritty cyberpunk for Shadowrun is easy enough. All styles are possible. "Ah, home."  gritty cyberpunk dive bar, busy crowd Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Version: v1.2.1 "Potion seller, I'm going into battle and I want your strongest potions"  gritty cyberpunk dive bar cyborg bartender Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 4, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Version: v1.2.1 These tools would also be excellent at things like padding out encounter table events. "I can see why I would need to make a dexterity check"  rope bridge over chasm Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Version: v1.2.1 "You want me to cross THAT??! There had better be an angry mob behind me with pitchforks and torches!"  pov standing in front of broken rope bridge over chasm, death trap decayed ancient neglected falling apart decrepit broken planks missing Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 2, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Version: v1.2.1 "Crap. Ok roll!"  angry mob behind me, pitchforks and torches Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 2, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Version: v1.2.1 Where I think things will truly start to shine is in some of the custom training options, such as making a specific character LoRA. You can take the custom artwork you have of a character and train a smaller model which you feed back into the generator which allows you to change specific things about the character, like outfits, action sequences, or battle wounds, etc. "That looks really bad, better take the Major healing potion and not the Minor."  Main Hero, bleeding battle wounds Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Version: v1.2.1 This is exciting stuff. I have to run and do some grocery shopping now, I'll be back later today to work on this some more. If anyone else has any other ideas I'm all ears. This will help me work on the GBS thread as well 
|
|
#
¿
Jun 13, 2023 19:10
|
|
|
Kestral posted:Something I keep coming back to about RPG use of these tools is how incredibly specific our image needs can be. So much of what we're trying to describe is, if not unique to our specific campaign, then so rare or specifically-detailed that it's impossible to find representations of it. https://www.youtube.com/watch?v=p52MHCcPc7Y 2:18 long. This is my favorite https://www.filmcow.com/ short. I don't know of any better way to describe the difference in specific complexity classes in so little time lol. I gave "angel whose many wings were made of old televisions and their spear is a TV aerial" a shot, and... missed. Here's an Imgur gallery of failures for you to see where things are with just word prompts alone. I usually make 4 pictures at a time to find candidates for further detail upscaling. Now don't get me wrong, this is a totally doable project, just not in one easy step like so many people keep saying. It takes work to get something so unique. Image generation training is based on word-image pairs, https://haveibeentrained.com/ is a site for searching through what Stable Diffusion was initially trained on. The most common picture of Angel will have feathery wings so by default putting in just "Angel" is most likely going to as well. There are methods for lowering probabilities like putting "feathers" in a negative prompt to tell it you don't want feathers in your picture at all, but that only goes so far before you need outside tools. One of those is called Inpainting. You can draw over a section of your picture and then regenerate it with a new prompt and it will fill in just that part with whatever you tell it to. I doubt even that would be enough for something as radically unique as television wings so we go even further into the many extensions available. ControlNet + OpenPose + Masking Tools + Krita Support You can take a child's scribble drawing and ControlNet will return a detailed version. So any level of drawing ability is enhanced by this. You can totally start with stick figures! There are also options for Posing a stick figure model and then your prompt will fit in whatever position you want it to. There are Depth Maps for making proper 3d effects. There are Masking tools for specific area selection such as just the wings for replacement. You could use all of the in Krita to take the initial feather wings and replace them with some kind of CRT effect texture. There will definitely be a write-up about ControlNet for the OP. It's a complete game changer. Now as has been mentioned since Photoshop first came out, you can totally make Anything. These tools can drastically reduce the amount of time required to make that Anything. You still have to put the effort in though. It's up to you to decide how much effort is worth it, a one-off low level encounter with a pack of goblins could be worth a simple generation where a campaign's BBEG would be totally worth using all the bells and whistles. Then you can train that result into a LoRA for easy repeated future use. I love your Bonsai tree analogy, mind if I steal that? Grey Hunter posted:Schrödinger's AI I also love that, mind if I steal that as well?
|
|
#
¿
Jun 14, 2023 06:44
|
|
|
Megaman's Jockstrap posted:Stable Diffusion Amazing writeup! Thank you for this! I just want to add there have been advances and an NVIDIA card isn't strictly required anymore. There is AMD, Intel, CPU, hell even iPhone support now. Yeah you can run this on your phone. Here is a list of quick installers, no messing with git or PATH required. - Automatic1111 - NVIDIA GPU - By far the most popular webui with an incredible array of options and extensions. https://github.com/EmpireMediaScience/A1111-Web-UI-Installer - InvokeAI - AMD or Intel GPU - This uses DirectML instead of CUDA. https://invoke-ai.github.io/InvokeAI/installation/010_INSTALL_AUTOMATED/ - NMKD - NVIDIA GPU - Executable windows GUI not webui. https://nmkd.itch.io/t2i-gui - ComfyUI - NVIDIA GPU - Node based with downloadable premade workflows. https://github.com/comfyanonymous/ComfyUI/releases - Stable Horde - NO GPU - Crowdsourced donated compute free for those without other means. https://stablehorde.net/ links to a client interface with no installation required https://dbzer0.itch.io/lucid-creations - Mac and iPhones - I don't know anything about Mac, here's a link anyways https://apps.apple.com/us/app/draw-things-ai-generation/id6444050820 - Runpod - 'The Cloud' - Just means someone elses computer. Rent someone elses GPU for money. Method of last resort. https://blog.runpod.io/stable-diffusion-ui-on-runpod/ - OpenVino - CPU mode - This takes forever but is still doable on literally a toaster, hell yeah stable toast. https://github.com/bes-dev/stable_diffusion.openvino
|
|
#
¿
Jun 15, 2023 05:12
|
|
|
Repaired Radio posted:One of my favorite parts of AI is that I don't have to pay artists if I want work done in their style. What's everyone favorite artists to try and replicate the style of using AI tools? My own.
|
|
#
¿
Jun 17, 2023 20:10
|
|
|
Kestral posted:Unfortunately, now it's going horribly wrong and I have no idea why! None of the techniques I've used up 'til now have been able to salvage that outstretched hand, which either becomes a gnarled lump of flesh and knuckles, or turns into faces. It's very haunted and I'm nearly at my wit's end. Ah well, time for more research. Thanks for the write-up, that'll be handy. It's late, I'll watch the video tomorrow and try to wrap my head around what could be going wrong, offhand I can't quite put my finger on it. Face it, it's hard enough to get ahead of these problems. I want to make sure I can help and not just give out second-hand information. Small tip though, lock in a seed when you find something you like. Much less randomization that way, helps to narrow down exactly what changes are being made by what setting.
|
|
#
¿
Jul 17, 2023 08:51
|
|
|
|
| # ¿ May 17, 2024 02:42 |
|
|
Kestral posted:People regard AD as nigh on miraculous, and you can watch it perform some incredible feats: It's funny, the same thing is happening in the video you linked. I went and gave this a shot since I need to learn this extension anyways, and I must say, this is quite a powerful tool here! It looks like everything could be done manually, but having the mask selection and Inpainting settings done for you is a huge time saver. Here's what I came up with and some  about what I found out. young tanned male artisan, white robes, grim determined expression, hand outstretched, burning savannah background scenery flaming horizon, lightning bolts storm clouds, 8k intricate detailed Negative prompt: nude, bare chest, open shirt, fleshpile, zombie, captcha, blurry, monochrome, greyscale, lineart, draft, minimalist Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 2, Size: 512x512, Model hash: 4199bcdd14, Model: revAnimated_v122, Denoising strength: 0.48, Version: v1.2.1, ADetailer model: mediapipe_face_full, ADetailer prompt: "(young:.6), (grim:.6), (determined:.6)", ADetailer negative prompt: "old, grizzled, happy", ADetailer confidence: 0.3, ADetailer mask min ratio: 0.001, ADetailer dilate/erode: 4, ADetailer mask blur: 8, ADetailer denoising strength: 0.32, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 96, ADetailer model 2nd: hand_yolov8n.pt, ADetailer prompt 2nd: open palm extended, ADetailer negative prompt 2nd: closed fist, ADetailer confidence 2nd: 0.3, ADetailer mask min ratio 2nd: 0.001, ADetailer dilate/erode 2nd: 4, ADetailer mask blur 2nd: 8, ADetailer denoising strength 2nd: 0.45, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 32, ADetailer use separate steps 2nd: True, ADetailer steps 2nd: 96, ADetailer version: 23.7.5, Hires upscale: 1.5, Hires steps: 32, Hires upscaler: 4x-UltraSharp - In the Detection settings, there's a dropdown menu with 'Detection model confidence threshold', that should be the setting to automatically determine which part get masked and then Inpainted. This can be adjusted per model, maybe that will help fix things? My guess is that 1st and 2nd model are getting crossed. - In Mask Preprocessing settings, there's an option for Merge, that too might be getting mixed up. Or why it's getting mixed up. Mask min ratio of around 0.01 should exclude the background according to the github page. - In Inpainting settings there's "Inpaint denoising strength" this determines how far from the original image you go when you regenerate it. It also is the fraction of steps used: i.e. .5 strength at 64 steps will actually take 32, .25 strength would take 16. Feel free to up the steps used at low denoising. A lot of experimentation is required to find sweet spots for 'same but different enough'. Inpaint mask blur will prevent the hard boundaries between masked area and what you regenerate, changing this can help blend everything seemlessly. - Oddly enough I couldn't get the hand detection to work, at all. This pictures hand is the result of the Hires fix. - Negative prompt is an incredibly useful tool. When you find a base image to start working with, add more terms to the negative prompt to get rid of the parts you don't want. Things like "bare chest" can help put that robe back on. Likewise for "woman" or "asian" if the model you're using is overtrained on those. Hopefully this helps some and thanks for bringing this extension up in the first place. I will definitely be using this in the future.
|
|
#
¿
Jul 18, 2023 01:54
|
|