Maximum Painterly Output!
This minor v1.1 update is nothing more than a merge with SAI's Offset LoRA. You can get identical results using version 1.0 and the offset LoRA at 1.0 strength. If you wish to control the effects of this more precisely, continue to use version 1.0 in conjunction with that at different strengths. Personally, I'm happy with this merge. The result is improved contrasts, and specifically provides much better night or dark scenes.
So you want to go all-in on a painterly effect? This model is not a comprehensive art style - it's a very specific painterly look. Alla prima, gestural, atmospheric. Detailed enough to warrant being called realist, but far from the clean sharp photorealism of Dutch or Renaissance work. Prompting for artist names only minimally pushes its output.
You don't need to prompt in any special way. If you want to use photography terms to influence the scene ('go-pro selfie', or 'by Max Rive') just start the prompt with 'painting' as well and you should still get a strong painterly image.
A great model for upscaling painting images:
Whether you generate an image initially with this model or another, if it is in an oil painting style, using this model in an upscaling workflow can help loads to keep that painterly material feel that too often otherwise disappears in upscaling.
If you feel compelled to offer some material form of thanks, you can buy me a coffee, or you can buy my (darn near unreadable, I warn you) short story 'Willem & Ellene'.
Description
simple merge with SAI's Offset LoRA to improve contrasts, especially to allow darker scenes
FAQ
Comments (22)
I love this! Have been able to get some fantastic paintings so really grateful for that.
The problem I am finding is that human anatomy is weird with this and I don't know if its the model itself but I find that it makes people with weird proportions like the head being way too big for the body for example.
Also when I try to make a woman riding a horse for example, I struggle to get the horse to not be tiny - maybe its a side effect of lots of portraits in the training or something?
the dwarfism is an issue that is more or less present in different iterations of training. Sorry. You can usually correct it with 'dwarfism' in the negative prompt.
I do know exactly where it comes from, but correcting it means removing a couple of images from my primary dataset which changes the style too much. Unfortunately, it's a couple images of kids and I don't really understand why these two particular images have such influence both positively and negatively but for whatever reason, they're part of the heart of this style so I'm living with the occasional disproportions.
Horse riding? I've had a couple of decent images, but I've always thought that was a challenge for SD in general. You find this particular model works against it in a stronger way?
Lastly, I extracted this (probably the last version of this checkpoint I will make) into a LoRA, as an updated v2.0 of ClassiPeintXL and, as a LoRA, can be more flexibly adjusted in strength and combined with other LoRAs so you may find it easier to work with and easier to mitigate tendencies that you dislike.
@eldritchadam Thank you, will try that when I am back to prompting.
As for the horse, maybe that issue is linked to the dwarfism but I am not sure. Even when I really heavily weigh the prompt to make a big horse or whatever it just often looks comical a lot of the time with the human to horse size proportion. To be fair, the issue really gave me a lot of trouble trying to prompt a kid riding a donkey as the donkey was much smaller than the child.
Not an expert at SDXL so definitely it may be just base sdxl limitations and my lack of skill...
I have experienced this 'dwarfism' too. Wonder what if you train with juggernautXL instead of sd_xl_base
@barepixels I'm aware of where it comes from. I can easily eliminate it altogether with the removal of two images from my primary dataset of paintings. They are in fact two of my own paintings of children and I was very ready to take them out, but the style changes dramatically without them.
Interestingly, I managed to tame the tendency in all of my latest Classipeint LoRA models, but when I packed those LoRAs into a checkpoint with some variation and additional LoRA models, the dwarfism returned. But it's not as prominent as in early iterations of Classipeint. And you really can eliminate it with one word in the negative prompt. So I'm not gonna fuss about it.
The final version of the Classipeint LoRA is actually an extraction of everything packed into this model. And the dwarfism rendering doesn't seem to be nearly the same issue with it. Results are also otherwise almost identical, so I prefer using the LoRA rather than this checkpoint, where I can modify its strength, combine more effectively with various models and LoRAs, and get a much broader range of output or, if I want to keep it simple, just use the base SDXL with the LoRA and get nearly the same result as if I used this checkpoint by itself.
@eldritchadam what one word in the negative prompt is that?
@barepixels dwarfism
I love generating art for my D&D game using classical painters like Caravaggio and the Flemish Masters. This checkpoint is wonderful.
Thanks so much!
I did the same for my friends' cards ^^
@diaserikcg114 I'd love to see some - yours too @cosophalas!
@eldritchadam I like to do things you wouldn't really expect with a lora, so I'll think about cool images. ;-)
I've used this model for a merge with RealCartoonXL v.6 and another painting model that was privately released by RaynaAI. It's just for personal use. Your model adds a little more coherence and depth to the texture of the brushstrokes. What I'd love to have is an unpruned version that is 13GB, and/or a Mega ART model that combines all your LORAs with this base model so that it can do just about any style or texture.
The only thing lacking now in the SDXL art model world is a model with thorough training on etching/aquatints/drypoints/engravings/woodcuts by Durer/Goya/Daumier/Hogarth/Rembrandt/Reubens, etc.
An ambitious-sounding project! I've not actually looked into model-merging with that level of sophistication. What's the right tool for that? Do you know of a walkthrough that could lay out the basics of doing such?
@eldritchadam Supermerger, which is an extension of A1111. I've tried to use it but can only do so for 1.5 models, as an XL model merger requires 64GB RAM using that tool.
I mostly use Invoke for all my generations as I like the output quality the most. and standard workflows for most of my ideas, and they don't have anything like that yet. They are trying to develop an XL merger tool, I've heard.
If you are interested in building a dataset for the monochrome model/Lora I mentioned, lmk. I've already been combing WikiCommons to find the highest resolution images (Tiffs and PNGs) from museum archives to build a collection.
Keep up the great work, and let me know if you have an unpruned version of this available. :D
@Crowyote alas, no unpruned version. My training is all poor-man's work on my lowly 8GB video card. The Painter's checkpoint is just a more simple mashup of my LoRAs into the SDXL checkpoint with Kohya, which is quite basic compared to the refinement Supermerger must allow. Maybe I'll look into remote computing on more robust machines, but before now I've always been primarily interested in seeing how much I can squeeze out from minimal resources.
Thanks for the encouraging words and the motivation - I'll look into it!
@eldritchadam Well that's pretty astounding and gives me hope for what is capable with only 8GB. Currently, you're ranked at number 3 on this site. You're bringing them lots of traffic. Not trying to go to your head or anything, but to but it in perspective. CivitAI has given away 4090 24GB card in some of their contests, seems like they should give them to the top 5 creators here . . .
I am a big fan of models like this that focus on non-photo, non-anime art.
I consider myself the anti-anime model trainer, LOL. Doesn't make my models the most popular, but I know you and I aren't the only ones looking for more than cartoons and nudity!
I downloaded this 2 months ago and it was pretty snappy. Suddenly it's bogged down my system and REALLY slow to create. I am using literally the same criteria as I was originally. Not sure what happened but whew.
That's a bummer! I wish I had an explanation, but I haven't made any updates to this model.
It's your GPU driver. Anything after 561.0+ on RTX, there's some kind of vram leak happening.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.