Hello! I am Sera/Seraluna. Recently, I have been trying to learn the intricacies of Stable Diffusion and LoRa creation. After experimenting with several different methods, this is my first results. My dream is to see more aesthetically pleasing LoRas to the community, and decided this is how I will contribute!

Thanks for checking it out! Read on for more information below. ↓

Documentation and advanced images coming soon!

About the Style:

This is my first ever LoRa I have created and my debut. This LoRa was inspired by cyberpunk anime/manga; think of the Cyberpunk 2077 comics. It was made with limited palettes, line art, and futurism at the forefront. Works best with character designs! I also find it to be pretty good at applying to dynamic backgrounds and city scenes.

About the Model:

Trigger word: Cyberpop (Optional)

The trigger word is not needed! But it does strengthen the LoRa style. (Example of LoRa strength!)

This model was trained using the super basic Anime SD 1.5 model. I have a lot to learn, so for now I was playing it safe since learning can be costly when training off my own equipment (I tried a few different training sources). But luckily I find it works amazing so far with most checkpoints I have been testing it on.

In total, there are over 9k steps! All high resolution and beautiful quality. The images used were mostly custom generated to fit my specific needs and tastes. I also included many diverse characters!

Using the Model:

(More testing needed. Remember, these checkpoints all have different needs. Parameters that work for one might not work for another. Play around with settings to find your best results!)

LoRa Strength: 0.7 (0.5 - 1)

Suggested Checkpoints:

• Your favorite one.

• Most anime checkpoints (Anime checkpoints work best since it's meant to be an anime style).

• Aurora

• AnyLoraCheckpoint

• Anything v5.0

• GodofSimp

Clip Skip: 2 (It was trained on 2, but 1 still works).

VAE:

• Your favorite one.

• Blessed2

Sampling: I have only tested Euler a and DPM++ 2M Karras, but both work good!

Sampling Steps: 25 (20 - 35)

HiRes. Fix: True

• Upscaler: Latent or R-ESRGAN 4x+ Anime6B (Only ones tested so far, but work great).

• Hires steps: 0

• Denoising strength: 0.55 - 0.7

• Upscale by: 1.5 - 2

CGF Scale: 6-11

Meant to be SFW, all images used to train are safe. Not tested with NSFW prompts or checkpoints. Might work, might not! No promises.

Prompting:

(Make sure to go over the Tags section below! It will help a lot here.)

Due to the nature of the tagging in training, I find simple prompts can be very strong with anime checkpoints. But it has been tested with more complex tagging on non-anime checkpoints, and still holds up fairly strong.

For negative prompting, the simpler the better all around! Sometimes I don't use any negitives. Just a few key words have been tested to work well (ex. watermark, worst quality, low quality, blurry, bad anatomy). Additionally, you can make more complex negatives if wanted and it does work super well with negative embeddings if you want more complex prompting.

Tags:

I curated the LoRa tags using Waifu Diffusion anime tags, then I manually tagged a lot of these images from there. Using these tags could help you achieve results closer to the intended style! Especially if you're using an anime based checkpoint, since a lot use Booru tags.

For more info and referencing, try using Danbooru tags or read up on Waifu Diffusion prompting.

Top 100+ Tags:

Cyberpop, cyberpunk, solo, 1girl, 1boy, dark skin, looking at viewer, short hair, jacket, black hair, shirt, simple background, male focus, long hair, outdoors, long sleeves, pants, holding, shoes, jewelry, standing, sitting, upper body, building, sky, earrings, scenery, weapon, bag, no humans, wide shot, hood, cloud, indoors, full body, smile, plant, shorts, window, blue eyes, gloves, socks, short, sleeves, skirt, mask, white shirt, black pants, closed mouth, hat, bandaid, tree, sword, red eyes, white hair, blue hair, shadow, tattoo, day, blue sky, multiple boys, ground vehicle, sneakers, backpack, from behind, teeth, belt, black jacket, blonde hair, white background, city, potted plant, katana, holding weapon, sign, sunglasses, open mouth, profile, hoodie, smoke, hair ornament, black footwear, headphones, red background, portrait, chair, black shirt, limited palette, horns, bandaid on face, cat, pink hair, flower, pink background, multiple girls, brown hair, open clothes, yellow eyes, cable, piercing, dreadlocks

Get the full list of tags in Stable Diffusion with the Civitai extension.

Suggestions, Feedback, and More:

Bugs or Issues: Like I said at the beginning, I am still in the process of learning. I would love to learn more about any issues that might occur while using this LoRa! If you do find a problem, and know how to solve it, I urge you to include that in the comment. This will help me learn and identify the issues faster.

Questions and Help: Feel free to leave a comment and I will try to get to it if I get time! However, due to my limited knowledge on the advances of Stable Diffusion, some questions I may not be able to answer yet. I encourage anyone with answers to help out!

Suggestions: Open to suggestions! If you are leaving technical suggestions on improving the model, and know how to do so, please try to include instructions to help me out. Also, I am open to hearing any LoRa model suggestions on what to make next!

Support:

Like what I do and want to support me more?! I am just starting off on this adventure, so join me!

Reactions 👍, Likes ❤, Downloads ⬇, Follows 🚶‍♀️, Tips 💰, and Positive Reviews ⭐⭐⭐⭐⭐ are easy ways to support me!

Twitter (X): I will be sharing images and little updates there.

Patreon: Still a work in progress, but will soon include early access, high priority suggestions, generated wallpapers, and more!