See full description and find LoRAs for other base models at this page: https://civarchive.com/models/971952
Update log:
(8/1/2025) v0.10 zero
Trained on 4.0zero. Synced latest 7K dataset and training parameters, close to illus v1.198.
(3/30/2025) v0.6 opt
Trained on 4.0opt.
(3/4/2025) v0.4
The plan was to bring RTX natural lighting and super detailed texture to 2D characters.
Trained on 4.0Zero with ~1k main stabilizer dataset, then with ~1k sub dataset contains real world images with dynamic lighting.
Unlike usual style/concept LoRAs, this LoRA was trained in a way that will only add small final details. It won't blow up your output and create a realistic 3D character. So the effect may not that obvious at first glance.
Recommended strength 0.5~1.
No trigger word needed. However, using keywords that related to some kind of light sources, such as "lamp", "daylight", "soft lighting", may bring better results.
In case someone may think this is placebo, here are 2 full size XY images. The composition of the images does not change. But the overall details change a lot.
It even generated diffuse reflection of the clothes on the table surface. (š bottom left image below)
Description
FAQ
Comments (5)
Would you consider training this on Animagine XL 4.0? The numbering is super confusing, but I think 4.0 gives the best images in terms of a good dynamic range of lighting, and the most creativity in terms of composition. I've been trying to switch my realistic model over to using it, as I started with Opt on v1. I've been "stabilizing" using very selective gentle merging of weights into various layers from a variety of checkpoints and speed/quality loras like dmd2. 4.0 makes fantastic images in 50 steps, but that is way more than people expect these days! I think the merging in Zero broke it a bit, and still didn't really improve hands or anatomy. Here are the original recipes:
4.0 = SDXL > pretrain 8M > pretrain 400k > finetune I > finetune II
4.0-Zero = pretrain 8M merged with pretrain 400k
4.0-Opt = Zero > finetune I > finetune II
Forgot to say - I've added this as a suggested resource on https://civitai.com/models/1378329/animagine-xl-40-realistic as it seems to have some effect, good work!
oops... this is a so fking late reply...sorry
I did train on 4.0, but I don't remember why I dropped it, iirc, it was the very strong default finetuned style that cannot be easily overlapped, also some brightness issues.
About hands, i think NoobAI solved this by training on many cosplay images. Not sure.
reakaakaskyĀ I see! I'm confused, you have trained in on zero? I much prefer the lighting on 4.0 rather than zero! I will give it a try though, thank you! :-)
ShinyLilly I trained on zero.
For my understanding, zero is pretrain model. Pretrain model represents the avg of the dataset and has no bias. It's for training a LoRA. So when you stack this LoRA on other finetuned model, you will get effect of this LoRA + finetuned effects.
If you train LoRA on finetuned model, because the finetuned model has high bias towards finetuned effects, which is not in your dataset, the model will have to reduce the bias to fit your dataset, which is not ideal.