Overview

A bunch of trained LoRAs for ChromaXL v3.31 (Mango Edition).

Apply with an activation keyword to significantly shift style in certain way. I usually attach an animation to showcase how LoRA transforms image after applying.

All (or, almost all) examples will feature only otters as maximally distanced from every possible copyright, patent or object-of-hate. Everybody loves cute otters, right? :)

PS: You can watch plans on training more styles in my google spreadsheets.

Styles / Activation

OIS - or, One-Image-Style

"Hey, I love how model made this image, I want to generate more images in this exact style, but if seed is changed, so is style... What do I do? Maybe, if I train desired style as LoRA... Well, I have only one picture that got usable, but... I guess one man IS an army."

Those are trained on exactly one image (usually, AI-generated), using LoHA (network dim = 8, alpha = 1, conv_dim = 8, conv_alpha = 1; noise_offset = 0.03, Cosine+0 warmup steps, min_snr_gamma = 1) + DoRA + Prodigy, for 800 steps.

• OIS / Plump Style - keyword digital drawing , produces a dulled-colors, 'inflated' creatures with big eyes and thick bodyparts

• OIS / Dark 3Drawing - keyword digital drawing , produces a darkened, render-like image where details is in more of 'modern' style. The less prompt you use, the more it look like a render, the more prompt (and weights) you use, the more stylized/'drawn' it becomes. The model might get weaker if prompted some kind of anime/smooth creature. However, this behavior is fixed in ISS.

• OIS / Watercolor Style - keyword watercolor painting , but you also can usedigital drawing for alternate effect. Initially, ChromaXL is unable to make a watercolor images, however with some elbow grease this style applies (or reminds, who knows...) just that - a feeling of painted, light imagery on a sheet of white paper. Notice: This OIS was generated on a free stock image of a overview of a forest and a lake, so you most likely to encounter a bodies of water if you use 'forest' prompt. Notice 2: If you use 'digital drawing' keyword instead of 'watercolor painting', then you will achieve nice alternate effect - you will still receive a fully-colored digital pictures, but with much less details, smoothed and shinier than usual.

• OIS / Pandora Style - keyword digital drawing , additionally you may use realistic too. Produces a somewhat realistic/photographic image. Might be somewhat blurry on the details, but nothing a good refiner won't handle.

• OIS / Fire Flower Style - keyword lineart. The effect is much more noticeable, if you use simple background (minor upgrade) or black background (major one) keywords.

SSS - or, Short-Stack-Style

"Hey, there is actually a bunch of images of similar style! The one image had small effect on getting it right, but what about exactly five?"

Short Stacks, as name suggests, are styles made on short datasets - composed of 5 images of the same-ish style. Usually, shortstacks are used to capture some artistic style in more detail than usual. They are trained for 2000 steps.

• SSS / Flowing Fractal Style - keyword digital drawing , to make effect more noticeable, you might also want to use abstract creature and <color> fur/feathers/body/markings to describe your perfect creature. Produces unique effect on, pretty much, any living species, and depending on a strength of LoRA, gives it either even more fluffier, or even more details. Experiments - encouraged. !!! This LoRA seems to be somewhat 'broken', as it inadequately reacts to weights ( red fur:0.53 and red fur:0.54 produce absolutely different images). Reasons: Currently unknown.

ISS - or, Iterative-Stack-Style

"Damn that OIS makes juicy pictures, but those aren't enough or aren't trained in specifics. I somehow need more. I need to refine it. I need... To train AI even further."

Iterative Stacks do have a lot in common with ShortStacks. They also have 5 images in their dataset. But the difference is - ISSes aren't based on artist styles, but actually on the previous iteration of OIS or SSS. So, you train a base model, you generate pictures using that model, and use generated pictures to train a new model! They share SSS'es 2000 training steps regime.

• ISS / Dark 3Drawing - keyword digital drawing . A direct upgrade over the OIS counterpart. I gathered and trained the most beautiful and most detailed images OIS had provided - with best fur, scales or feathers there is. As a result, this model is capable of drawing near perfect fur, feathers and scales on anything, really. If OIS image would look like a flat, Source Engine style render, then ISS version would look as professional Unreal Engine XXXXL Hyper Puper Duper style render.

GISS - or, Grande-Iterative-Stack-Style

GISS is an ISS, that was trained on 25 images, that was generated by OSS, SSS or ISS. But they were all generated with the same prompt... Why? You'll know, if you read below (soon enough).

Grande Iterative Stacks infinitely lot in common with Iteratives. But this time, Grandes have a VERY SPECIFIC dataset - a 25 carefully picked up, balanced prompts, that was generated by a lot of seeds from OSS/SSS/ISS source. This not only smoothes the style out, but actually improves coherency. Downsides? You get ~3-7% less style per GISS iteration. A small price to pay for massively increased stability and opened way to create RHMSes.

Those are trained at ~2800 steps.

• GISS / Plump Style - keyword digital drawing . A direct upgrade over the OIS counterpart. Downside - less plumpy, and things like fur may have took a hit. Upside - much, MUCH clearer, not as fuzzy as original, and less prone to bad generation.

• GISS / Dark 3D Drawing - keyword digital drawing . Something of both an upgrade over the ISS counterpart and a bit of downgrade on the other hand. So, on the upside, the style became clearer, consistent, and sharp. On the downside, it is not 'the same' style - the fur became less visible, less 'realistic'. Understandable, since I first trained an original image into OIS, then OIS into ISS (losing 10% of style), then ISS onto GISS (losing additional 10%).