/hdg/ Prompt Assist
Links to other Collaboration Edition Guides/Resources Home LoRA Training | LoRA Repo
This is a non-exhaustive list of ways to generate good images with your prompts. It'll include links to additional resources as well as few examples provided by myself. Please experiment! This is the only way to move this medium forward! Happy Prompting!
Why does my Generation look so desaturated?
Apply a VAE in settings manually, auto doesn't work properly (as of 2023/01/03) you can find additional VAEs within hdgrecipies.
Why does my Generation randomly turn into a green/black square?
This is due to your GPU not supporting
half precision floating point numbers this may be fixed one day but currently the solution is to right click your
WebUI-user.bat file and under
set COMMANDLINE_ARGS= add
--precision autocast -- no-half-vae this will result in more VRAM usage but should stop the black box generations. See StableDiffusionWebUI Troubleshooting for more information.
Where did all of the LoRAs go?
They have been moved to a seperate Rentry to make space! Please visit LoRa Repo
Links to additional community guides sdgfaq | hdgfaq | hdgrecipies | NAI-Speedrun
- /hdg/ Prompt Assist
- Character Tags
- My Usual Prompt Structure
- Tag Ideas - /hdgpa/ Edition
- Schizo Negative
- Some of my Prompts
- Other Anon's Prompts
- This is what you want the AI create. Put whatever you can think of in here, the AI will then proceed to try and create what you've written as the output.
- As per Anon, tokens are read from Left - Right 0 has the highest strength while 75 has the lowest, resetting upon hitting token 76 and so on. Disclaimer: I don't know if this is actually true, Anon says it's within the Code though so maybe.
- NEGATIVE PROMPT:
- This is what you want the AI to avoid. Put whatever you DO NOT want in here. The AI will then proceed to create something without the tags you've put in here.
- SAMPLING STEPS:
- This tells the AI how long it should spend on the image. The Longer the better generally but stick try and stick to lower as to save time between generations. I've personally found a comfortable range around 16-32 steps.
- Sampling Method:
- This instructs the AI on how to think about your Image. Different Methods provide different results for the most part in testing you might find matching results.
- This is how you tell the AI the resolution of your image. If you go too big you might have a VRAM Memory error. (CUDA Error)
- CFG Scale:
- This tells the AI how "focused" you'd like it to be on your prompt. Lower value = "less focused" higher value = more.
- Source number for the beginning of AI Processing. Two images with the same seed SHOULD result in the same image depending on the version of WebUI you are using it may differ slightly.
- In basic terms; this is a miniature model/hypernetwork that can be hotswapped(?). These can have some great effects on generations.
- Currently these aren't a standard feature within WebUI and require an extension which you can get here.
- We don't currently have a repository for these yet but I may look into doing that if LoRA Anon's don't do it first.
MORE INFORMATION CAN BE FOUND HERE: NAI-SPEEDRUN txt2img training wheels I used information from this guide for this one as well.
This link brings you to another rentry with information on tags you can use within your prompt as well as pictures showing displaying the effects of the tag.
NovelAI Anime Girl Prompt Guide
Despite saying NovelAI in the title you can use this for any model that's been mixed with NAI without any issues. It has tags with pictures accompanying them to display the effects of each tag. It also provides some larger example prompts to base your first prompt off of.
Danbooru Tag Wiki
Most models were trained on images found within the booru, you can use any of the tags found within this wiki and see some sort of effect on your generation. Mix and Match!
This Torrent has a majority of the main models and mixes everyone is currently using, download the ones you want and seed! If you can download them all please do! We can always use more Seeders!
AbyssOrange_mixes_fixes Magnet Link
This Torrent has models that have been mixed with AbyssOrangeMix, this was trained on Instagram photos so it should have a decent sense of anatomy.
This is a bit older and not up-to-date so I'd suggest checking the top two first but this is also a place where you can get models!
Google Drive Mirror
I made this, models will be uploaded slowly (and when I have free time) because they take forever.
AOM2 Mixes MEGA | Magnet
Concepts/Poses | Characters | DiscordEmbedDump
Embeds are essentially prompts within a file, you call upon the file within your prompt and it will use the keywords encoded within the embed to generate the image it is associated with.
UnNamed Anon's Le Malin
What the fuck is a LoRA?
- Fine-tune Stable diffusion models twice as faster than dreambooth method, by Low-rank Adaptation
Basically, it's a faster Dreambooth/Hypernetwork. Less work, better effect. (in most cases)
Having used them for some time now I'm fairly confident in saying you don't need style tags in your gens with a LoRA involved.
LoRA Training | LoRA Issues | LoRA Repo
How to use a LoRA
- Install this extension.
- In your main WebUI in the bottom left extend the window for "Additional Networks".
- Check the "Enabled" Box.
- Under the Model dropdown select the LoRA you want to use. By default these are located in \Stable-Diffusion-WebUI\extensions\sd-webui-additional-networks\models\lora you can change this in WebUI's settings panel.
- Adjust your strength to your liking.
- Generate an Image.
--medvram fix has been implemented by Kohya, you don't need to do anything besides update now.
If your LoRA stops working after a Model Switch run a prompt at a different weight then go back to the weight you want.
We've moved them all to their own Rentry to save space, all of them (that we've previously repo'd) are there though! Please check them out!
As you can see, this can have some CRAZY good results! Prompt
My Usual Prompt Structure
These are personal preference, there is no one right way to do this. I just like to have structure.
- Decide Whether or not the Image is gonna be NSFW, place the NSFW tag at the front if so.
- Pick out other style tags. Almost always recommend:
- (extremely detailed CG unity 8k wallpaper)
- (best quality)
- (best illustration)
- (best shadow)
You can also add or swap out for:
- (oil painting)
- Decide the main focus of the image. What angle/composition are you attempting to generate?
- Go into the details of the character you are generating. (Hair Colour, Eye Color, Breast Size, Extra Features [Fox Ears, etc.], Clothing)
- Decide on a background/location or keep it simple.
- Detail the Pose, Facial Expression, etc.
If you're using a LoRA skip step 2
I definitely didn't adjust this to turn it into a Yandere Prompt...
Tag Ideas - /hdgpa/ Edition
Work in Progress! Please send ideas!
Dakimakura combined with a FINAL resolution of 512x1856. Example from Anon
Arm up, arm raised, (armpit:1.4), presenting armpit, (angry:1.1) SOURCE
Curvy Curvy, Plump, Wide Hips Combo - Meido-Anon
Short|Medium|Long (Color) Hair
Short worked for me at one point not sure if it's a constant though.
You can get very detailed with clothing, see here for an example.
Occasionally you might read this term on the board... it's exactly what you think it is:
This isn't required for good gens, but it definitely works.
Some of my Prompts
Anzu, The Blonde Kitsune
This prompt is short and simple but generates what I believe to be a really great image. My favourite comment being "It was as if she was looking at living garbage" and it's true I am trash.
Cornelia, The Noirette Yandere
Other Anon's Prompts
Thanks to everyone in the /hdg/ threads and anyone who's guides or images I've linked! Continue pushing this medium forward! Happy Prompting!