///DISCLAIMER

I AM NOT A PROFESSIONAL. THIS ARTICLE CONTAINS MY OBSERVATIONS ON THE 200+ LoRAs, MAINLY CHARACTERS IN PONY AND ILLUSTRIOUS, PUBLICLY AVAILABLE FOR DOWNLOAD AND ONLINE GENERATION. THIS APPROACH HEAVILY DEPENDS ON TIME, EFFORT, AND YOUR VISUAL CAPABILITIES UNLIKE MOST TRAINING GUIDES WHICH HEAVILY RELIES ON THE POWER OF YOUR COMPUTER. THE OBSERVATIONS MADE MAY OR MAY NOT BE OF HELP IN CREATING LoRAs UNDER OTHER BASE MODELS BUT THE APPROACH PRESENTED CAN HELP YOU CREATE MULTIPLE CHARACTERS, OR MULTIPLE CONCEPTS IN A SINGLE LoRA. ///

PROLOGUE

Hi and Hello 😀

My obsession with fanarts and an easy mental therapy solution led me to creating LoRAs XD, how about you? Why do you want to create a LoRA? Is it to create fanarts and spread the existence of a character? To sustain your degenerate and deviant urges? Or to provide for the Creativity–helping inspire others to do what is hoped for them? 

ANALOGY LIST #1:

1. Checkpoint / Base model = Your Color. 

2. Steps / Step count = Your artistic strokes

3. CFG Scale = Your Influence

4. LoRA = A box containing your Idea or Inspiration

OBSERVATION LOG #1: A single checkpoint/Base model is not God

• Did I think that I can create any character from a single checkpoint? 

  • Yes, before discovering CIVITAI, I was fiddling around at NightCafe and LeonardoAI. I wanted the AI to generate a specific scene with specific characters and elements—naturally a failure. 

• Did I try different approaches? 

  • Yes, in prompting: from simple layman language to narrative, and even using emojis/alien letters from a language not known by most–but known by the ai. It was useless but I concluded with how different checkpoints have different “brains” on the type of prompt you are using

  • Yes, in the UI: I concluded how steps between 30-60 is the goldilocks zone for checkpoint models. Depending on what checkpoint model you are using, they will have their own favored/desired step count. The vision/end-goal appeal you are trying to achieve must also influence what checkpoint you are using. Especially on checkpoint merges, You cannot force your opinion on specialists. Base Model and Checkpoint Trained are what I prefer to use especially when wanting to mix juxtapose concepts

ANALOGY LIST #1 EXPOUNDED:

1. Colors(checkpoint / base models)= you can write / draw almost anything with a greyscale pigment / color but need to consider some risks when using different colors. Force your idea to a red ink and some people might perceive it as a sign of aggression or passion. Just like checkpoint merges who can easily create peaceful settings with a serene style, forcing the idea of war will create an odd image. By odd, you might generate a group of adults simply bickering among each other under a vibrant open grass field and not blood and gore with dramatic lighting.

1. Artistic Strokes (step count)= A traditional painter generally does not need to scratch the paper like a chicken, or a writer does not need more than three words to create a compelling story. But the more strokes, the more details are added. More details is always appreciated but, someone with a vision and understanding knows how much is enough steps and too much will create unnecessary noise

1. Influence (CFG Scale)= How much influence do you want your prompt to affect the generated output? Or How influential do you want your words to come out? Flux is a great model who will follow your prompts 95% of the time. Know that it is recommended to follow the CFG scale mentioned in the models’ page. You don’t want to be super confident with your words that you want everyone to believe in you 100%. A machine can easily believe in you, yes, and it will give you A LOT of noise ON THE FIRST FEW WORDS. Notably on models not built like a skyscraper like flux and SD3.5 which have higher capabilities in what they know. 

   1. TIP # 1 = Try to limit your words to less than five if you want to use CFG over 20. Flux can handle over 100 ← I could only test up to 100 with TENSORART’s limits but am guessing it can >.>. CFG heavily depends on the capability of a checkpoint model→how beautiful and easily controllable their grayscale shades can create. The more fine-tuned a model, the more you’d want to stick with CFG less than 12.

1. Box of Idea or Inspiration (LoRA)= You like something←an element from whatever you are referencing, and you want that specific element to be easily generated without using much tag sequences. You need to train / package that idea into a box / LoRa.

CHAPTER 1: HOW DOES ONE CREATE A LORA?

You now know your goals on why you want to create a LoRA. For me…

Goals:

• To Prompt the Character Likeness in any Outfit and Scenario

• To Maintain Character Likeness Integrity

• To Attain a Flexible LoRA

Now, how does one create a LoRA?

OBSERVATION LOG#2: It's EXHAUSTING BUT WORTH-IT APPROACH

• A LoRA needs a dataset. A dataset contains images and text files with similar file names which contain tags that describe the elements of the image being referenced.

• Why is I’s approach more exhausting than simply taking a picture you like and tagging them properly? 

  • My approach is cut-cost but big effort. 

  • Ask the questions…

• Where to collect images and tags?

• Who is the main focus of the LoRA?

• What kind of images should be collected?

• When should some tags be used?

• How should the images be tagged?

• What available base model in CIVIT’s on-site training is best for the dataset?

• Requires you to visualize an average generated output after going over all the consistent elements in the dataset. 

• control over consistent elements in all images. Elements consistent in 5 images should be taken into consideration whether they should be in a unique tag or remain as is.

• Control over tag sequences in the txt files. 

• Control over the amount of different kind of images

• Control over the visual composition

• I have observed how tagging everything in a dataset and if found consistent in 5 or more images can greatly help in removing/adding that specific element, and the elements around it.

CHAPTER 2: A DATASET FORMAT

A Dataset Format is how you structure your dataset. The number of each kind of images in different aspect ratio, and how tags are present or absent in txt files. 

OBSERVATION LOG #3: The General Ideas

1. Different Aspect Ratio can help add a natural dynamic angle to your generated images. It most definitely helps you include images that are…

• In dutch angle

• With negative space to the sides (left, right, above, below)

• Cropped images

1. Each image contains a single or many elements and there should be corresponding tags for each of those elements otherwise, the elements around a tagged element will associate themselves to the tagged element. Example:

Tags Associating the elements present are

pokemon (creature), leaf, blue sky, cloudy sky,

1. The higher the quality of an image the better. Be vigilant for consistent elements that you do not like

2. Images with a quality lower than 480p should be tagged with the schizo bad negatives, “blurry, bad artist, bad art, low quality, etc.”

3. The autotagger can easily describe elements in high quality images

4. Tagging allows the association of a consistent element and the elements around it. The more images containing visual elements and their descriptive tags the better a checkpoint model can remove or add those elements.

5. 5-20 images with a consistent element is enough to be recognized by a base model checkpoint. Recognition is important so that a single element will not be bulldozed by another. 

6. Describing elements should be simple and general. The autotagger is your best assistant at the same time your betrayer if you do not properly check your dataset tag frequency. 

7. A single image with an untagged element and is consistent on 15-30 images will commonly stick to the images generated

8. Only uniquely tag elements that you wish to add or remove

There is no LoRA if you do not have a dataset. Different kinds of LoRAs need different dataset forms. Aiming for a style and thinking how training will expand on the images you provide will not work. The training process will not train itself and certainly not generate really good AI images. As stated in Chapter 1, a dataset needs images and each of the images corresponding text files.

Where to Collect Images?

• Collect the images either on the internet (very efficient), 

• prompt them (training efficient–because AI-Generated images somehow are faster during training processing + You can improve on your prompting skills),

• manually draw / photo manipulate for the idea (Workaholic efficient).

—-------------

TLDR:

Training a LoRA for Pony

Concerning Pony Images:

1. I recommend 60-90 images, minimum of 30, maximum of 350.

2. Do not include any images that does not have elements or barely have any resemblance to the focus of your LoRA

3. Aspect ratio should be either at a width or height or both of 1024x. You can go for 512 to reduce the cost further but you will feel the need for upscalers

Concerning Pony Tagging:

1. You may or may not use the autotagger

2. Tag every element in the image

3. Focus your unique tags around what your LoRA is supposed to generate 

4. Give 1-3 unique trigger tags.

5. The number of tags should vary and not have the same count on all images except for the 1-3 trigger tags

6. Blurry and low quality images should be tagged with tags you use in the negative prompts

7. Do not bother adding quality tags such as scores, masterpiece, etc.

8. Prune repetitive tags into one unique or general tag

Training for an ILXL LoRA

Concerning ILXL Images:

1. I recommend 30-70 images,minimum of 1, maximum of 1000. At 120+ images, the average style is replicated even without the use of hidden tags.

2. Do not include any images that does not have elements or barely have any resemblance to the focus of your LoRA

3. Aspect ratio should be either at a width or height or both of 1024x. You can go for 512 to reduce the cost further but you will feel the need for upscalers

4. ILXL is better with characters and concepts surrounding characters so you may or may not feel the need to add extra images

Concerning ILXL Tagging:

1. You may or may not use the autotagger

2. Tag every element in the image

3. Focus your unique tags around what your LoRA is supposed to generate 

4. Give 1-3 unique trigger tags.

5. The number of tags should vary and not have the same count on all images except for the 1-3 trigger tags

6. Blurry and low quality images should be tagged with tags you use in the negative prompts

7. Do not bother adding quality tags such as scores, masterpiece, etc.

8. Prune repetitive tags into one unique or general tag

—-------------

OBSERVATION LOG #4: Current Dataset Forms

FORM 1: When creating Characters…

• Try to go beyond 30 images for a character. Why?

  • For the character’s likeness to be recognized. 

  • For the character LoRA to have a wider capability range in the prompting playground. 

• 30% of the total images should be close-ups of the character’s face and hair. As much as possible, avoid including any other elements besides their biological characteristics in this 30%. 

• The rest of the 70% can either be medium shots (shots from the torso up), or cowboy shots.(shots from the knees up)

• One full body is enough, unless the character has other full body shots with clear quality. Always include them

• 3-5 images of different perspectives/angels the characters are viewed from is enough. Perspectives/angles are either the character being viewed from the side, from above, from below, or from behind. From the side and From behind are important perspectives, 

  • If there are no such available perspectives, resort to cropping

• 3-5 Extreme close-ups are a must for unique details. These can either be the shape of the character’s pupils, or their forehead, or a strand of differently colored hair. I sometimes forget to add their extreme close-ups–see the LoRA “Brant” or “Zilong” where the differently colored hair strands are unstable across the showcase.

• If an element is deformed, be sure to tag them with the appropriate negative prompts

• Don’t bother about the background or the lighting when creating a character LoRA. But, if the background or lighting heavily influences the normal appearance of the character, appropriate descriptions such as “dim lighting, sunburst, or landscape, from afar, etc.” should be added.

• If the character is well known by the autotagger, consult the base model or any of your preferred checkpoints and try prompting for the same character. Check if a character’s physical elements difficult to remove or add

  • If it is difficult, change the autotagger’s tag for that character into a unique tag

  • If it is not, keep the autotagger’s tag for that character

SUBCHAPTER 1: Tagging

STEPS:

1. Input Core Tags (Main Tags to trigger LoRA effects)

1. Subject Core Tag: Who is in the image? (ex. IOC)

2. Subject Focus Tag: What are they? (ex. cute blueberry-round character focus)

   1. Are there 2 recognizable subjects in the image? Yes

      1. If hetero use (couple focus)

      2. If homo uses (yaoi focus) or (yuri focus)

   2. Are there more than 2 recognizable subjects in the image? Yes

      1. If family I uses (family focus)

      2. If group of friends I uses (group focus)

      3. If a crowd/a number of people I uses (crowd focus)

NOTE:

I advise to describe the focus carefully as the word before ___focus, will greatly affect the output.

1. Is it a non-human or humanoid? Yes (ex. pokemon (creature))

OBSERVATION:

Pokemon (creature) tag is a highly versatile tag. Though I think that the tag only contains creatures from the pokemon series, its versatility can help in focusing the LoRA on whatever shape/form the subject has.

1. What is their gender? (ex. 1other)

   1. If couple, (ex. 1boy, 1girl)

   2. If homo couple, (ex. 2boys) or (ex. 2girls)

   3. If non-human (ex.1other) or, describe the closest animal/thing to that subject (1blueberry creature)

NOTE:

I am careful when describing something/someone’s gender. They do not discriminate but once there is a clash of information from the subject focus, training will go haywire and output LoRA will not listen.

1. Is a consistent background present on more than 5 images? Yes (ex. crossover)

1. Input Other Unique Tags

   1. Does the auto tagger not describe an element that is perceived? Yes

      1. Creating a Unique Tag following the structure

         1. Color (ex. Blue-green)

         2. Subject Core Tag (IOC)

         3. Element being described (skin)

blue-green_IOC_skin

1. You may or may not add subjective tags like “happy, atmospheric perspective, depth of field, etc.”

SUBCHAPTER 1 Questions:

Q1: Why three or more Core Tags?

• I have found how three core tags increases LoRA flexibility and power, allowing the LoRA to take effect once there are two or more core tags present. One core tag may be added at a higher prompt order while the other can be at the lowest prompt order while still remaining to procure the LoRA’s effects. This is most effective when using multiple LoRAs or adding different characters into the prompt structure.

Q2: When to use the - symbol?

• I love to use the - symbol to bleed (merge) concepts together. I thought of it when they observed how the autotagger adds “fur-trimmed” or “long-sleeved” to a single element that can be described with two tags.

• If I see an element have two features that can be easily recognized by the naked eye, I describe first the feature which they think has a big impression-followed by the second feature. (ex. A blue haired boy with yellow strands of hair = blue-yellow_male hair)

Q3: When to use the _ symbol?

• Space (   ) and underscore (___) are different.

OBSERVATION:

• using space on UNIQUE TAGS can confuse the checkpoint model

• Using ___ to connect types of tags into 1 tag.

• It sounds similar to using the - symbol but the - symbol is used to describe a hybrid element while the ___ is used to create a unique tag

Q4: What about a Limited Dataset? 

If some characters have either only one or two images, I try to achieve the bare minimum of 15 images by adding extra images or simply cropping. These can be naked body types that have a similar frame/appearance to the character or duplicating the characters’ images and individually crop the elements to their own image. Their clothes, and limbs included. I always add these extra images—-NOTING THE BASE MODEL.

Some base models are not welcoming to duplicating images which is why I prefer creating character LoRAs for pony, and more so with Illustrious. Then generating more datasets with checkpoint trained base models such as obsession ILXL. You can refer to DaBerry’s LoRA which I will continue to add more dataset for its retrain to a better and flexible LoRA

Q5: Why Illustrious?

It is super easy to start creating a character LoRA and concepts with Illustrious

Though it does not have a wider general scope as pony have, its specialty allows for easy anime-styled fanarts and any concepts related with cartoon / anime style

FORM 2: When creating Backgrounds, Buildings, and Vehicles…

• I always try to go beyond 40 images for inanimate objects or settings

• The images must be either in a day, afternoon, or night setting

• I do not include humans in the images unless it is a background. 

• There should be less than 10 images with humans

• I does not concern with the perspectives and only focus on what the average default appearance would be generated in the Prompting Playground 

• Too many repeating elements–though varying–will still result in Element Anchoring. I believe that repeated varying elements that are not recognized by the autotagger but seen clearly by the naked eye is worse than concept bleeding. It is an element that anchors to a tag, it may or may have anchored to more than one tag and using those specific tags will result in that element showing up. Why is this worse? There will be a sussy amongus event in finding those sussy tags

SUBCHAPTER 2: Clutter Tagging

• How do you counter Element Anchoring? Instead of going against it, I decided to flow with it, using a unique tag but also not a unique tag–clutter. 

• They discovered how it can add a little bit of noise, and when there is a little addition of details there can be a blooming foundation! So I decided to make categories with clutter

  • 1stclutter is for main elements in the scene

  • 2ndclutter is for the secondary elements

  • 3rdclutter or Mclutter is for man-made clutter. These are typically smaller details of the image

  • 3rdclutter or Nclutter is for natural-made clutter. These are typically plants/nature

  • 4thclutter is for anything lighting related…(ex. sunlight, shadow, reflection, etc.)

FORM 3: When creating Compositions, Poses, Clothing

• 30+ images for concepts

• The bare minimum goal is to transfer the subject to the target concept, pose, or clothing.

• There should be multiple subjects in that concept

• I struggle with style bleeding from the concept images as of the moment but they have observed how form 3 is best delivered with a dataset containing images in multiple styles. 

SUBCHAPTER 3: Focus Tagging

• I only prunes the tags describing the concept/pose/clothing

• If the concept or an element is not recognized by the autotagger, I duly add the unique tag/s concerning those elements.

  • When tagging with compositions, I will for sure add unique tags to describe the composition. For example, (insert H comp example)

  • When tagging with poses, I will check for the gestures which Autotagger recognizes and prunes them for a unique tag. For example, (insert Dasit pose ex)

  • When tagging with clothing, I will remove all clothing related tags and add the unique tags necessary that describe the clothing. (insert Daflow of cloth)

:: HIDDEN TAGS ::

• I have observed how “anime screencap” denotes generating images with the anime style—true, but only true if paired with an anime style or the tag “anime” with it. Otherwise, the tag itself is similar to the clutter tag. I use it to anchor the average style of the entire dataset and hope to reduce its influence in the training phase. I thought of it before they were fiddling around with network dims and network alphave XD

• “Crossover” tag is another hidden tag which I thought was useless other than the autotagger describing two characters it knows but those characters are from different shows. After playing around with it, and hypothesizing with delusions. I think that the crossover tag helps remove a consistent background. Because if it were to create two characters from different shows in the same scene, the background should be most affected.

• If wanting to blend weird concepts with a human/living being e.g. (a concept of a person holding a clock for a face or an animal with half their body in different places) “surreal” tag is best. Use as a hidden tag, otherwise if prompted, will introduce weirder concepts that can really be considered as surreal. The tag can even cause inanimate objects to merge with what you do not expect or biological placements such as eyes to go astray.

• “Style parody” and “parody” dictates an artstyle following an anime screencap style. The autotagger does not / rarely give out the anime coloring and more so the anime screencap tag so the first layer of tags to associate the consistent style is “style parody and parody” you can use this as an alternative to anime screencap but I prefer to use these tags in a tag sequence for fluidly changing the artstyle of a character

• Negative space is a photography tag describing the blank space between the main focus and another element. The autotagger does not give out this tag and I only discovered it with its effective function in backgrounds. It opens the AI for proper distancing between elements. This tag is not useful if the main focus of the prompt is multiple characters.

• “Expressive face” enhances the prompt adherence to emojis and > < keyboard letters tags. It can also turn eyes to dollar signs or apply itome.

• “focus” a tag for focusing on the main constituent element of a LoRA. It can also be applied for increasing prompt adherence on prompting with non-LoRAs

• “Monster girl”, “species” and “monster boy” naturally stretch your character's tags with the non-human tags you are trying to merge them with. Similar to “surreal” these two tags are double edged tags and you are better off using pokemon (creature)

• Genderswap (mtf) or genderswap (ftm) works well if you place it together with the opposite gender tags. So male focus → female focus, and 1boy → 1girl = VIce Versa, you can easily genderswap anyone with said tag sequence. In training, it can be used as an alternative if wishing to merge two different looking characters into 1 main core tag but differentiated by the presence / absence of “genderswap (mtf) or genderswap (ftm)”

• Lastly is the elephant in the pony–or room >.>, “pokemon (creature)”. Absolute madness, this tag is. It is similar as to how pony was supposed to be made for generating ponies but ends up being a general core tag for anything that can be considered as “creature”. WIth the tag, I have experimented how even inanimate objects can be given a humanoid trait, maintain subject integrity, and differentiate the subject from other elements based on how repeated that subject is across the dataset. Drawback is when the LoRA is prompted at a lower weight where an actual pokemon (creature) becomes a photobomber

• I do not know anything deep in the details of each training parameter. I simply follow a cheap estimate of 1060-1075 total steps..

Pony Training Parameter for Dataset Form 1

Pony Training Parameter for Dataset Form 2

Pony Training Parameter for Dataset Form 3

Illustrious Training Parameter for Dataset Form 1

Illustrious Training Parameter for Dataset Form 3

Where is Illustrious Form 2? 

I believe in the focus of a base model. Observed how Illustrious is best made with characters and concepts but unsuccessful compared to Pony when creating Form 2 LoRAs.

OBSERVATION:

• I like to keep the total steps around 1060-1075

• # of Epochs is set to 1

  • I have also experimented with higher epochs and think that, yes, it can improve the LoRA’s capabilities→only so the elements would be associated well by general tags (ex. Shirt, long hair, dress, etc.)  ← this is a I delusion, DRINK A CUP OF SALT FOR THIS ONE

  • Approaching a higher epoch with current dataset forms and processing will result in a null. A similar result or maybe worse than having 1 Epoch. 

• Shuffle tag is checked cause, why not :///. It is considerate enough to think of future I’s promptings.

• Flip augmentation is useful for symmetrical designs. I only used it once→the twins from Ouran High Host Club

You may or may not see me as a cheapskate but in fact I only want to create fanarts efficiently and freely. They never wanted to create LoRAs or spend much existing-time rotting in hardwork. CIVITAI’s on-site training allows the former’s life to flourish. Achieving Character Likeness while maintaining its integrity for a basic flexible LoRA. Yes, basic curiosity cannot be sated and wish to expand their capabilities but, at present, only bound to what CIVITAI has to offer. Maybe in TENSORART? It is still growing as well. Nonetheless, I am happy to create and exert efficient hard work to share the respectable hard work of others.

D. DICTIONARY GUIDE

Question #1: What is a LoRA?

• A helpful and guiding resource for the checkpoint to focus on. It contains a dataset. Visualize any ingredient, they have their own functions. Each LoRA has their own approach in guiding the checkpoint.

• Another imaginary example is: Visualize a person, they have their own characteristics and personalities. Each individual has their own approach in guiding you to your goal. 

Question #2: What is a dataset?

• It is a group of images and notepads. The images show what the checkpoint should generate.The notepads/txt.files tells when the checkpoint should generate. Visualize images as the nutrients of your ingredient, and visualize notepads as the timing of when these nutrients are at their peak deliciousness.

• Another imaginary example is: Visualize images as the characteristics of a person and visualize notepads as the action of when their characteristics shine best. For example, you know someone with an average normal facial expression and ticklish. So you tickle them, and they react appropriately with a beautiful smile.

Observation #1: Descriptions are tags/captions

• Notepads contain descriptions for a representing image.

• Descriptions can be either structured into tags or captions.

• Tags and captions have the same function–literally. Their structure is how they differentiate

  • Tags are short words example (man, woman, fruit)

  • Captions are a sequence of words example (a man is watching, a woman is reaching out her hand, a fruit on the table is fidgeting with fear in its eyes)

SideQuestion #1: WHERE to use Tags and WHERE to use Captions?

• First and foremost, check/remix on the showcase images of the checkpoints in the preferred base model you will be training the LoRA on. Ex.(Pony base model = AnimeboysXL)

After Checking:

• If the positive and negative prompt follow tags, similarly follow the same structure.

• If the positive and negative prompt follow captions, similarly follow the same structure.

(Further asking why will engage you with a I delusion)

Question #3: What is on-site LoRA Training?

• It is where files or a ZIP file is sent for a website’s on-site training.

SideQuestion #2: What is LoRA Training?

• Visualize an apprentice learning how to hone their craft. They repeat their days without any optimal practicing, and learning, until they become masters.

• How they train are the training parameters→repeats, epochs, steps, etc.

  • Repeats = The count on how many times an image will be repeated/observed. Visualize a classroom and the teacher keeps repeating a lesson because each student needs to learn the lesson at their own pace

• Epoch = Visualize an hour of a clock as a single epoch, what you do in that hour are the repeats. The second hour would be a second epoch, what changes here are the amount of steps being utilized.

• Steps = The energy cap the trainer needs to learn. Higher steps can lead to overtraining or what they say, overbaking. Visualize a highly competent worker, they work so hard they decide to quit the day after. Similar to an overbaked LoRA output, they try to generate with your prompt but they end up discolored or plain static. On the other hand, lower steps can lead to undertraining or salmonella. A half-assed painting is not going to sell, similarly with an undertrained LoRA, the checkpoint cannot give you a quality piece with that LoRA because it has not been fully trained.

• Network Dim / Network Alpha = I do not know and simply could guess. Default CIVIT settings works enough..

OBSERVATION:

• I believe that their character LoRAs can be improved with a better understanding of training parameters. But since I am a non-material lover of this wonderful world, they dislike understanding the technicalities of a machine. So they try to learn which tags the machine is competent with. Then try to estimate/imagine how the output will appear with what a dataset form contains.

You made it this far? Congratulations! Thank you so much for taking the time to read. I’m just an art enthusiast looking for ways to bridge my homeworld to this world. EW CRINGE RIGHT???? But, I hate their current standing and a world that you hate is not where your home is. I wish to create the phantasia, and after much experience and knowing the boundaries of what they are currently capable of, it will now be sooner applied than later thought. Huhu, am just happy to have this opportunity and I truly and genuinely bid you dear reader, a good day and adieu~