An evolution of Envy Daybreak, this produces sharp, detailed digital paintings, but in a more versatile color palette. It's also good at hyperrealism and photography.
v1.0+vae is the same as v1.0, but with the VAE embedded for convenience.
Nova and Aurora aurora more opinionated about style, and don't replace 1.0.
Galaxy is a model that focuses on highly detailed digital paintings.
Quasar is even more detailed than galaxy, but uses a more muted palette.
All versions are compatible with Starlight DoRAs.
See posts below for example images using the various Starlight DoRAs.
Description
FAQ
Comments (14)
Can you update your posts so we can see the prompts? Having a hard time using the model.
That's because I made them in comfy. If you drag one of them onto the screen in comfyui, it'll load all of the nodes.
I'll post some more samples from auto1111, though.
Looks very interesting! I'm a fan of many of your 1.5 models so i'm excited to try this one!!
Edit 1: Alright, I don't know if this is accidental but, when I started testing this model i forgot to change my prompt (it was for photorealism) and it is producing astounding photorealistic results. It really surprised me, I will be posting some examples soon.
Edit 2: Also, it works supper great with IP Adapter and InstantID, this might become one of my top checkpoints for the kind of work I do! 👍
I'd definitely be interested in seeing those photos. :)
And yes, I also love IP Adapter, particularly the style-only mode on ComfyUI. For whatever reason, it does seem to generate better results with it than most other models. No idea why that is.
Could you please share some details on how this model was trained? What style of captions did you use? tags or natural language? Does it have other models in the mix?
Also, what Lightning model did you merge, the 4 step or the 8 step?
It's complicated. I don't remember precisely what all went into it, but I do know that it was basically Dreamshaper XL Lightning merged with the difference of Envy Daybreak XL 02 and (I think) vanilla SDXL .
Rather than standard finetuning, the way I like to make adjustments to my models is by training sliders (https://github.com/sdbds/sliders-for-windows) that increase or decrease a particular characteristic, and then do some testing and merge them in at the best weight possible. I generally train those sliders themselves on vanilla SDXL, because it has a relatively broad understanding of styles.
My testing has indicated that the optimum number of steps is around 7, but ymmv.
Hi, I've been playing with this checkpoint, and I really like the results, comes out nice and crisp with a cool aesthetic. There is one thing that bothers me though. My ladies seem to frequently come out with very long torso and sometimes double decker boobs. Have you come across this as well? Do you recommend square aspect ratio, is there a good portrait ratio to prevent this, or is there another fix?
I generate a bit smaller and then scale up a bit more. Try keeping your original height to 1024px or less.
Love the model ! I'm having a hard time with the hands, it loves to add fingers, shorten some, longer some and twist them in all direction. Any hints or tips ?
I realized recently with my DoRA training that if I'm not careful to include training images with hands, Starlight will basically forget how to draw nice hands. In short, it may be associated with a DoRA that you're using, and if it's any of mine (which is probably is), let me know and I'll see if I can re-train them for better hand performance.
Also, make sure you're using at least 7 steps, DPM++ SDE & SGM Uniform, and also running HiRes fix if possible (2-4 steps at 0.2-0.4 is enough).
@_Envy_ Thanks ! I'll try the things you propose. Been using the Fae dora a lot lately, will go back to base model and see if I have differences without any DoRA,
All your Starlight DORAs are tagged "SDXL 1.0" not "SDXL Lightning". Is there a non-Lightning Starlight checkpoint that I'm not seeing, or are all the DORAs just technically mistagged? (I'm not saying you need to change the info in all the DORAs, I'm just making sure I'm not missing something.)