shadow scroll
Kyaa! Just the other day, I stumbled upon a post by Kaiser-senpai https://discord.com/channels/1122227993805336617/1172995113199341668 , and it was all about using the blind box method with MyShell's image generation bot to create some super kawaii pictures and then make your selections. But, oh no! That method tends to eat up a lot of power, so I'm here to share a more energy-saving strategy that can most likely help you get those top-notch, sugoi images.
worldly geyser
Can it really be controlled so precisely? Amazing!
drowsy wagon
good idea!
shadow scroll
New discovery, you can just copy and paste it
oh forgot to add something more, once you get a template from gpt4, you can copy paste the template and ask gpt4 to generate another high-quality article about anything you want with fewer and casual keywords
rugged creek
thank u Evmil!
atomic lotus
thank u Eeeeeevmil!
shadow scroll
thank u Evmil!
happy if it can help u
shadow scroll
thank u Eeeeeevmil!
happy if it can help u!
silk ermine
The idea of using MyShell's image generation bot along with an energy-saving strategy is brilliant. It's always exciting to discover new ways to create amazing pictures. I'm eagerly looking forward to more of your sharings and the wonderful images you'll create!
gaunt void
Hi there, I'm new on the server but noticed there was some interest into image generation using a the prompt system of Stable Diffusion (the one used by CivitAi).
I noticed there was an attempt to generate a picture with a narrated description based on the tags used of a picture.
I'm not sure but perhaps that was done to save some tokens?
Well, there's some details that might be of help while dealing with SD:
Common prompt: 1girl, blonde hair, dress, ribbon, tie, blue eyes
Weights added: (1girl:1.3), (blonde hair, blue eyes:1.4), (dress, ribbon, tie:1.3)
The usage of weights help to greatly emphasize certain details of a picture, thus, if you had several "regens" with the eyes always getting a different color, adding a weight can increase the priority of the tool to use that color. The values can go from :0.1 to 2 if you want, but less than 1 (1~0.1) means less chances of something to appear and more than 1 (1~2) increases the probability by a lot. In general you won't use more than :1.2 ~ 1.5.
With all that being said, saving tokens comes from the reduction of punctuation to process text in a more straightforward maner which, although do helps with the token usage, halves the chances of SD producing a picture with the previous mentioned advantages of it's system.
At the end what could be said is that: if the image you wish to generate uses few tags, straight up use those tags and even add the negative Prompts. If the picture uses more tags, then reduction might be needed but at the risk of losing definition.
That's all!
lapis fjord
wow man, thanks for the head up , keep the good work going bro
shadow scroll
Hi there, I'm new on the server but noticed there was some interest into image g...
woooow it's my first time receiving such a warm comment, thank you!
i'm not a technical creator so just shared my experiences in using image bots of myshell, i'm not sure what the "Weight" would mean since never tried it
and the reason why i only used long paras to generate images is that their image bot is called "natural language image generation" so i thought i just needed to write something for images, and indeed it worked!
for the negative prompt, yes i also heard from other creators but it seems the bots there don't support it or im not using it correctly
im quite new to the image generation so just could share my experiences based on my interactions with them, which might be different from technical usages, (i often saw discussions about SD but haven't try it on my own yet)
overall, thanks again for your detailed and warm comment! love your patience and care!
gaunt void
No issue! I'm glad to be of help
Now to answer some of your questions:
"Weight" would be the value I mentioned before. That ":1.4" I placed after the "tag". A weight helps in the sense that the more value (1.4 ~ 1.6), the more said characteristic will take relevance in the image generated. For example, if I use a prompt and include "hat", the hat might appear in any size. If I use (hat:1.2), then there will be a hat of average size. If I use (hat:1.4 - 1.6) I'll get a very detailed and probably big hat. You can do that with pretty much any "tag" to get better generations.
On the other hand, there's the negative imput. By using (hat:0.8), you'll decrease the size of the hat. At (hat:0.6) or less there will be no hat at all since it's now removed. That's a way to add "negative imputs". Another way to do so is using "[ ]". A tag placed between those will be treated oppositely as those with "( )", which means if I use [hat:1.2], then now hats will be smaller and at [hat:1.6 ~ 2], hats will be completely banned.
If you use prompts like "bad anatomy", "extra fingers" or "Low quality" as negative prompts (using either the "( )" or "[ ]" method), then you will exponentially improve your generated pictures. Sadly, this also means the usage of those symbols for SD to understand them (which means more tokens for a LLM). So yes, it's actually pretty hard to be "greedy" with tokens while using the standard SD image generation procedure.
As a side note, the "Loras" are "extensions" of a model. For example, you mentioned "Meinamix", which is a "model" or "checkpoint merged" (just call it model, don't make your life harder 
) while you used the lora "Lora:100starsA-1:0.6", which is obviously a Lora. To explain what's the difference, let's say the model is the cake and the Loras are the cover. The cake (model) contains almost everything you need and can happily eat it as it is, but the cover (Lora) adds a layer of personality to the whole dessert to your specific taste. A more practical usage is the Ai-generated pictures of anime characters that we see everyday. Creating a "model" takes a lot and a whole lot more of pictures to feed the AI. On the other hand, creating a "Lora" of a certain character uses fewer pictures but MUST be of the character in question. Using this method you can replicate art styles, characters, even objects or clothes patterns.
In relation with the previous point, if you don't have a way to "add" a Lora, then there's no need to even mention it on your prompt since SD won't know what to do with it. On the other hand, if you do find a way to "include" the Lora in the model, then a new set of rules appear: remember the weights mentioned before? Well for Loras you have to work with values below "1". ALWAYS check the description of Loras since they contain keywords that will trigger their appearance in your generated picture (for example: "Scarf", "Raiden Shogun", "Armor", etc) AND, more importantly, the recommended weight to use them. Most Loras use ":0.6 - 0.8 " since, just like the cover of a cake, you just want to add a layer of sweetness, not a whole damn layer of sugar. Using "1" or more for a Lora can exponentially ruin your picture since you'll be overdosing the poor model. There's no need to explain how that works since at this point this has gotten pretty confusing probably, but all you need to know is that :0.6 ~ :0.8 are the safest for Loras (still check the description).
I hope this explanation can be of help, and if it wasn't, well I tried 
kindred flowerBOT
dznalientsu has been warned
brankopro6396 has been warned
brankopro6396 has been warned
kindred flowerBOT
hypemethepro8902516 has been warned
kindred flowerBOT
brankopro6396 has been warned
kindred flowerBOT
smolkakyoin5374 has been warned
kindred flowerBOT
opanda7060 has been warned
kindred flowerBOT
smolkakyoin5374 has been warned
kindred flowerBOT
opanda7060 has been warned
