RaynaEnvision XL is a model named after me [Rayna] that generates images that have a "painterly" feel to them. I have an art degree that I never use, and I miss making art, so this is kind of a way to do that without quitting my job.
UPDATE: An improved version, RaynaIlluminate XL, is up on TensorArt. It can do goldleaf! It's also much better at getting Gustav Klimt right. You can also download it on my Ko-Fi page.
My first checkpoint was designed for realism (it's only up on TensorArt, but you can play with it there) so I thought I'd go the opposite way this time.
Description
FAQ
Comments (17)
It has a good wide mix of color outputs and randomness of colors between prompts
Amazing, thanks for making this!
I'm also an art graduate and am excited to try this <3
Yes! Excited to see what you come up with.
Ah, the comments on my image review did not appear for some reason.
A+ for effort. The CFG resistance is pretty high, I usually use 27.5 with a complex prompt, progressive sampler (heun most often) and a higher step count using the NMKD GUI (1.12 Invoke backend) Img2img with no skip clip, no refiner and no LORAs. In some generations, I got extra limbs and rubbery features. Compared to most checkpoints the painterly effects were more consistent from seed to seed in a batch, but not as deep as some others can acheive texture-wise. The colors were also much better in this model than others as the palette is much more muted and consistent like an oil painter that builds an image with just a few colors mixed.
What I would like to see in version 2 is more realistic painting/brushstroke texture, maybe pastel imitation, etching, engraving textural effect. This checkpoint gets closer than most of the other XL ones, however, it needs deeper training on various styles and painterly effects. Wiki Commons is full of high resolution images from collections of the NGA, MET, and Cleveland museum etc of classic artworks. This could be the definitive model for those of us trying to avoid pure photorealism, pixar stuff, etc.
This is a really good and thorough review -- the exact kind of thing I like to hear about, so I can learn and improve. Thank you so much for all of this input.
@RaynaPolsky You're welcome. I hope development is continuing and that you feel encouraged. I have changed my step counts recently and really work from 10 > 125 or (250) : 1 ie. 10 through 250 steps and putting out each single image to study . My CFG ranges now from 7.5 to 27.5. Heun is still my default because it is probably the most cohesive progressive sampler to build an image. I have tested your model with some LORAs in NMKD 1.13 and got decent results but coherence and consistency in unruly resolutions like 2048 x 768 can suffer.
I would strongly suggest studying the RealitiesEdgeXLAnime20 and RealCartoonv2 models for their ability to stay coherent in almost any resolution and to have more impact on overall dispersion of details. The palette of your model is better that these overall, however, if the texture and consistency can be improved this would be the prizetaking art model. I wish you well!
@Crowyote I actually created a whole new model that works much better, called RaynaIlluminate, that I'm currently putting up for sale on Patreon. I'd love to give you a free download link, if you'd like it, in return for all this wonderful input. Are you on Discord? If you send me a message on there I can give you the link.
@RaynaPolsky Aww thanks, Rayna! I really appreciate that. I'm also really glad to hear that my input has been helpful. I'm not the usual kinda SD user. As one art degree holder to another we gotta help each other out and fight this tidal wave of waifus abuse! 🤣
Love the results, can this be used on sd 1.5 base model?
RaynaEnvisionXL is a base model -- Civitai also refers to them as "checkpoints." You might be thinking of a LoRA, which is a kind of model that can be used alongside base models.
do you have a list of artists this model responds to?
It should respond to most artists that SDXL knows -- I was surprised there were so many of them, actually. Give it a test run and experiment -- any list I could give you would be thousands of names :)
Spectacular. Thank you! Congrats.
Beautiful samples, can tell you love the art
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.