DOWNLOAD NEGATIVE EMBEDDING HERE FOR BEST PERFORMANCE.
This is Aether Lux, a Lexica-inspired 2.1 model, trained on synthetic data. Images are preferably generated with a base side of 768px, and for general purpose, Euler works well. For photorealism: DPM++ SDE Karras is the best choice in my experience.
The model is accompanied with a negative embedding, lux_np. You need this one applied to the negative prompt to make sure that the model functions at its very best. You place the file in the Embeddings folder within your Stable Diffusion installation (usually stable-diffusion-webui\embeddings).
Description
FAQ
Comments (8)
While obvious to most, it might be helpful for newbies to include a comment about where to place the negative embedding file (\stable-diffusion-webui\embeddings)
Thanks! Sure I will do that.
hey, out of pure curiosity, what exactly is the negative embedding doing here and whats was your approach on making it? If i had to shoot a guess, im guessing you're taking some of the bad results or even training sample images and feeding these into a negative embedding so it knows not to do this mistakes its known to have? something like that, sorry - have seen people do this a few times and just finally thought id ask what the approach is.
Usually a negative embedding, as with any ti embedding, is trained on a specific set of images generated with the same model it’s destined for. In my case I just tested out different negative words or phrases in the negative prompt that I later merged into an embedding just from that. Training one is more efficient though, but I was in a bit of a hurry and I feel this neg embedding works well. People can and should still add their own things in both prompts to get precisely closer to what they want.
This model is just amazing, and dare I say: much better than Lexica. The colours and contrast is just much better. Love it!
Thanks! 🙂
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.