Roto_1 Is my attempt to develop a style that achieves a rotoscoping effect, while avoiding a flat-color cel shading effect. While I am generally pleased with the result of this first outing; It produces subjects with realistic proportions that have an overlay of outline and inline work. However, there are some caveats:
Prompt Adhesion: The higher the weights, the lower the adhesion. I think this might be due in part to my dataset which may be too small to realize my vision. I recommend not having the weight set any higher than 1.0
Filter Bypass: After some experimentation. If you typically use "Vanilla" Krea2. I recommend using a filer bypass (Krea2FilterBypass is what I used), This has more to do with prompt adhesion than it does with NSFW generation.
LoRA Compatibilty: After some trials, it does produce the linework I am expecting to see even when paired with other LoRA, but you will have to tool around with varying weights to get the sweet spot.
Trigger: I made a trigger: Rot0, which I don't really think is necessary. Try it out, if you wish, but the LoRA's effect seem strong enough without it. Seriously, as seen in the comparison images, Roto_1 varies wildly from the original prompt result.
The Dataset
I prepared a dataset of roughly 100 mages using Sick Ollie (which is one of my current faves for photorealistic images) To each subject generated, I added the following style prompts:
rotoscope‑based rendering, natural proportions, expressive motion capture energy
style: neo‑rotoscoping, 2.5D dimensional shading, DC comic linework overlay, hand‑inked, 90s theatrical cel aesthetic
rendering: dual‑layer look (ink line layer + painted shading layer), smooth gradient cel colors, subtle painterly textures, controlled halftone accents, non‑photorealistic realism
animation: drawn animation still, movie‑quality rotoscope finish, crisp silhouette, dynamic pose clarity
background: detailed painted backdrop, atmospheric depth, soft volumetric lighting, negative space composition
color: rich but controlled palette, warm shadows, cool highlights, gentle color falloff
camera: medium shot, stable framing, cinematic composition
quality: high resolution, clean edges, refined shading transitions, production‑grade finish
tags: rotoscope, 2.5d shading, dc comics style, cel shading, ink linework, animation still, gradient shading, painterly texture, atmospheric background
Once created. I used the CiviAI LoRA trainer at their default setting, running 16 epochs. This LoRA is the 12th, which met my expectations.
Overall, I had fun working on this LoRA, and I hope you find some value in it before moving on to the next "biggest and best". Thanks, for giving a Promptist like me a chance 😊
Description
FAQ
Comments (1)
It kind of struggles to get the linework, but I am unironically loving this at lower weights (<0.8) for adding more organic shadows and lighting to scenes where I'd have to previously specify them. It doesn't change composition so drastically at those weights either.