This is a fine tuning of SD1.5 and designed as a general purpose model trained on high quality photographs and traditional artworks upscaled and denoised before training to get the sharpest cleanest results possible. It's trained at 576 resolution using the offset noise fix so generations are sharp and detailed with vibrant colors, deep blacks, and a well balanced contrast.
This model was not trained with A.I. generated images or merged with any other models which means your images won't have that green/aqua color cast seen in so many models today.
The largest portion of the dataset consists of photographs of women, men, gorgeous landscapes, and luxurious home/cabin/hotel interiors, some abandoned buildings and cityscapes, followed by a few unique art styles that can most easily be discovered by prompting with the commonly used words below.
Some tips for realistic images: avoid using the word "realistic," part of the dataset contains "realistic porcelain dolls" and the weights for "realistic" have shifted in the direction of the dolls. It shouldn't be an issue but If your faces look a bit smoother than you like try throwing the word "doll" in the negative prompt. This model is great at both high detailed realism and stylized images when prompted correctly.
A few commonly used words found in the training dataset include:
Papercut
Liquid Splash
Realistic Porcelain Doll
Interior
Landscape
...and a few others for you to discover.
If your so inclined to leave a tip I'm happy to accept Monero at this address: 82s3fk8bQB2DHJ3r9idZUsST1Dvf5cKKC6Fu87rYgV9dAFbCbAcXMPXaP59yDwWzRXfYfTBszHZno6xGwDb17xUzEkCsAah
HuggingFace: https://huggingface.co/Cosmo-Hug/FeverDream
Thanks for checking out my model, feedback and advise for future updates are appreciated, enjoy!
Description
FAQ
Comments (11)
Vae baked in?
Darnit, I really wanted to try this one. 😔
I'll upload it without VAE :) Give me 15 minutes
@Hartman that's what I baked into it. Just figured it would simplify the process of trying to remember if a model needed a vae or not and which one. Out of curiosity, why wouldn't you use the model if the vae is baked in? I didn't foresee that being a problem for anyone but I'm sure there is something i'm missing.
@Cosmo88 some LORAs don't play well with models that have VAE baked in. Lots of artifacts, glitch pixels, distorted faces.
And in rare situations, changing the VAE doesn't work.
@wktra I see, I don't mess around with LORAs very much so thanks for the heads up!
@Hartman Not an issue if you're using auto1111 but some don't, and there are more tools out there that don't support VAE than do, at this point.
@imn0mn When using Auto1111, some LORAs produce junk artifacts when paired with a model that has a VAE baked in. (Even if that VAE is switched off or changed the settings)
@Cosmo88
As an artist, I thank you for uploading the non-baked Vae version.
@wktra Not suggesting there's no use case for a model without baked in VAE. Simply asserting that there's a use case for a model WITH baked in VAE as well.
@imn0mn I'm not criticizing you. I just want more creators to be informed. That's why I explained it.
make a lora of this please, and maybe a seperated one especially for the color splash effect too?