Victorian Anime Art XL版本VARXL (ID: 412695)

This model was painstakingly trained in the course of a week on 500+ images of mostly images I generated with Noisy_Element/VAA/VALite, as well as the dataset from the Victorian Anime Art LoRA. Couldn't get it to train on Colab, due to out of memory issues, so I had no choice but to go the slow method.

This model, like Noisy_Element_XL was trained on 4 categories: full body, Victorian, Monet (the painter, not me), and close up, thus those are the trigger words to invoke a few things. SDXL models and I haven't gotten along well in the past since it struggled to do full-body shots, so I tried to rectify it here. The trigger words are of course optional, as it should work fine without them.

Just like Noisy_Element_XL, this model is baked with the SDXL VAE, thus no external VAEs are necessary.

Recommended Usage:

Run as normal (non-LCM samplers, CFG at 7, steps at around 30)

Though this is essentially the non-LCM version of Noisy_Element_XL, it is compatible with LCM. If you wish to use the LCM samplers on this model, simply use the LCM LoRA in your prompts as you use this checkpoint. In the case of using the LCM LoRA and Sampler, remember to reduce the CFG to about 1-2, reduce the steps to about 8, and you should be good to go.

SDXL LCM LoRA: https://huggingface.co/latent-consistency/lcm-lora-sdxl/resolve/main/pytorch_lora_weights.safetensors?download=true

描述:

Victorian Anime Revolution XL (couldn't come up with a better name, lol)

This is essentially the non-LCM version of Noisy_Element_XL_v2.5. In order for this to resemble its LCM counterpart, I had to tweak a few settings, since simply removing the LCM LoRA from NEXL v2.5 by itself would only lead to a noisy mess. Even now, VARXL seems to get noisy if you set weights above 1.3, so be careful of that. From what I can tell so far though, if you do things right, this model comes out cleaner than NEXL v2.5.

Just like its LCM counterpart, this checkpoint is merged with some of my current LoRAs; Victorian Anime Art XL, STFD, and RealisticV (a prototype style LoRA which definitely needs work, but I suspect is what causes the art style of this checkpoint).

Theoretically, this model should be LCM compatible--provided that you use the SDXL LCM LoRA with it. In addition, this model is baked with the SDXL VAE, so no VAE needed. If, for whatever reason, you get an error message trying to use this checkpoint, then try manually selecting a VAE of your choosing to go with it.

训练词语: