→ about

a new version of Luna Diffusion with more realism. is it better? probably subjective
painterly style is less strong, but details like eyes and overall anatomy have improved, & has more contrast, sharpness
skin texture is also stronger, which can be a good or a bad thing depending on your preferences (personally i've had enough of airbrushed skin so i like it) and your setup -> for example, output with the a1111 web ui looks softer than the one made with the vladmandic fork. might require more adjustments with cfg scale or prompts than the previous version depending on that
use this VAE if your images come out fried
also checking "Enable quantization in K samplers for sharper and cleaner results. This may change existing seeds. Requires restart to apply." in Settings -> Stable Diffusion in the Web UI can help

finetuned on a few hundred & mostly hand-captioned highres images on SD 1.5
no trigger words/ tokens, but you can add "painting" to the prompt to get the painterly effect
use "illustration" in prompts to get more vector art looking images
works best at 768x768 px, 512x768 px or 768x512 px since it was finetuned on 768x768, so 512x512 will look overbaked
DPM++ 2M looks usually nice and crisp, use Euler_a for a more softer look. UniPC seems somewhere inbetween
i recommend adding “nude, naked” to your negative prompt if you don’t like boobas because this model certainly does (￢‿￢ )
check my older blog entry for more examples of the previous version, comparisons and tips on settings

🧨 Diffusers

This model can be used just like any other Stable Diffusion model. For more information, please have a look at the Stable Diffusion Pipeline.

from diffusers import StableDiffusionPipeline
import torch

model_id = "proximasanfinetuning/luna-diffusion-v2.5"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "painting of a beautiful woman with red hair, 8k, high quality"
image = pipe(prompt, height=768, width=768).images[0]

image.save("./result.jpg")

you can also get it as Safetensors

→ some comparisons

→ example settings on A1111 web ui

prompt: "painting, portrait of a beautiful woman, soft lighting, surrounded by flowers, urban city background, high quality, black purple iridescent hair"
sampler: UniPC
steps: 12
CFG scale: 6
768 x 768
VAE: vae-ft-mse-840000-ema-pruned.ckpt

if you enjoy this consider buying me a coffee or becoming a monthly supporter

(ノ◕ヮ◕)ノ*:・゚✧

license

This model is licensed under a modified CreativeML OpenRAIL-M license.

Utilizing and hosting the Luna Diffusion model and its derivatives on platforms that earn, will earn, or plan to earn revenue or donations requires prior authorization. To request permission, please email proximasan@protonmail.com.
You are permitted to host the model card and files on both commercial and non-commercial websites, apps, etc. as long as you properly credit the model by stating its full name and providing a link to the model card (https://huggingface.co/proximasanfinetuning/luna-diffusion-v2.5), without performing any actual inference or finetuning.
The Luna Diffusion model and its derivatives can be hosted on non-commercial websites, apps, etc. as long as no revenue or donations are received. Proper credit must be given by stating the full model name and including a link to the model card (https://huggingface.co/proximasanfinetuning/luna-diffusion-v2.5).
The outputs of the model or its derivatives can be used for commercial purposes as long as the usage is limited to teams of 10 or fewer individuals.
You can't use the model to deliberately produce nor share illegal or harmful outputs or content
The authors claims no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in the license
You may re-distribute the weights. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the modified CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully) Please read the full license here: https://huggingface.co/proximasanfinetuning/luna-diffusion-v2.5/blob/main/luna_diffusion_license.txt