Introducing the Next Evolution of Quillworks

This marks my first major overhaul of Quillworks, built from the ground up using a completely new training methodology. With a fresh dataset of over 12,000 carefully curated images and a higher network dimension, this model is designed to deliver richer details, improved color fidelity, and greater variety in its outputs. It represents a clear departure from the original version; so much so that I’ve released it as a separate model entirely.

That said, this version is still a work in progress. The rapid shift in training strategy introduced some unexpected challenges, especially in the tagging system. As a result, outputs can occasionally be inconsistent or unpredictable (hands in particular are a known weak point right now due to style overlaps). Improving the tagging is my top priority moving forward.

Even with these hurdles, I’m incredibly excited to share this model. The gains in texture, shading, and visual depth are already substantial, and I believe this version showcases the direction I want to take Quillworks in the future. I hope it inspires you as much as it motivates me—there’s a lot more to come, and I’m just getting started.

Usage Guide

CFG: 3-8 have been tested. Lower output seems to turn the images into more of a painting and higher outputs seem to prefer anime. Around 5 you get an interesting crossover.

Scheduler: I highly recommend DPM++ 2M using Karras for experimental pieces. However Euler a using Simple, also gives interesting output that seems better for hands.

Steps: 30+, I run most of my images at 35. This is crucial as it allows the model to add the fine details necessary for the textures to appear properly.

Prompting: For best results always include a style at the start.

Positive Prompt Example: "Manga sketch, colorful, 1girl, geisha, close up, beautiful, vivid eyes, "

Negative Prompt Example: "Worst Quality, bad, awful, "

But this model does more than just manga sketch styles, feel free to experiment with a wide variety of options like "Oil painting, Brushwork, Sketch, Watercolor, anime, Cartoon, Digital painting, Low Poly 3d, Semi realism 3d, and a few others."

This model runs well at normal illustrious/SDXL sizes. If you want larger, I recommend using a hires fix like R-ESGRAN but the textures like brushstrokes and sketch lines may not fully apply due to the hires fix. Hires is often a great way to fix your hands or eyes.

Please remember the output on this model is MUCH more experimental than I did on my last version, do not be surprised if this model does weird things, has broken images, or outputs nonsense!

As of now this model is focused on characters with a bit of landscapes. Landscapes are currently more limited in style and a plan for a future release is to improve backgrounds and landscapes.

Open Source: this model was made by merging illustrious2.0 and hassakuv2.2 with both having my trained Lora applied at strengths of .8 and .75 respectively. The Lora I trained was run in Kohya for 63 hours to complete 2 epochs. While this is a merge, it has a massive amount of new data added to give entirely unique output with a special focus on texture and color.

final note, I added a style called "absurdism" and it does REALLY weird stuff.