L1N3S

New I_pomo mix models. Three models. Cool illustrations. Check on civitai for more information at civitai.com/models/652317.

The objective with this new version was to get very difficult illustration prompts right, and in many cases even without upscaling or without post edition of any type. A lot of methods were used to get to that objective that cost me a lot of work and experiments, and I think for now it is satisfying enough. Most images won't need upscaling. Sometimes the model will distort faces, so there is still an experimental vibe to the model, but this is uncommon. Good prompting from the user side is still needed, so I recommend an LLM to translate that ultra-cool illustration look that you want into a proper prompt, or you could just use my examples; all example images have the workflow included; just import it into comfy. I recommend that you use 48GB VRAM for quick inference, but it works fine with 24GB VRAM (lots of OOM errors with 24GB VRAM can happen if the illustration is too complex). The model this time is bigger, 1 GB, but it is worth it for the quality we get in the end. I will try in future versions to train smaller sizes while attempting to keep the high quality of outputs. The style tends towards abstract with dynamic brush strokes, paint splatters, etc.

Features:

• This Lora was trained to be the most accurate on very difficult illustration prompts. The eyes, even at distances, keep sharp even in the middle of the chaos. So you don't need to upscale, but in various cases the use of upscaling or injection is nice and it is recommended.

• Ipomo basic art with enhancements, some mix of Pony magic, NAI lighting, a mix of Niji wonderful colors, compositions, and variations, and heavy cool illustrations. This defines L1n3s, the new iteration of the previous I Pomo anime mix.

• A newly carefully selected dataset of more than 183 images, with the objective to keep the anime with maximum quality while keeping the illustrations really heavy. Some experimental techniques were used to get these results directly in the Lora in order to avoid upscales and other methods (the example images are all using these methods).

That said, illustration loras are the hardest for mixing with anime; they simply tend to blur everything and ruin the image, so I still recommend that you use every single method available to make it better (upscale, injection, etc.; check the workflows in the example images).

Use l1n3s for maximum variation and I_pomo girl as the trigger words. Both need to be in the prompt. The best initial weight is 0.85.

Trigger words

You should use l1n3s to trigger the image generation. I_pomo girl needs to be in the prompt.

The different versions

The original keeps the style closer to the original I_pomo Mix.

The filmic Niji version has the colors and lighting of Niji and a filmic effect added that increases the the color complexity (the face gets more saturated, but the colors are better) and the Niji style towards a darker comic style is in there, with the little colored traces and bright one lines around faces, that people love (reminds of the 3D effect when you look at a picture without the 3d glasses). It also changes the original style of the face a little.

This version is experimental. More tests are needed to validate this model. I ask that you guys test it if possible to help development.

The Ink+ has more pronunced Niji style with all the cool touches and illustrations and colors from the Filmic version. The faces blend more with the background in heavy illustrations.