the🤦2.1 update
v2.0 was baked with the SAI offset LoRA - this made the model interact poorly with fine-tuned checkpoints also utilizing the same. But even with the base SDXL model, lended a warm color cast to outputs. All-in-all, 2.0 was flawed. 2.1 is all that 2.0 was supposed to be, with the SAI offset LoRA stripped.
Version 2.0 after working on a full checkpoint model, Painter's Checkpoint, I felt I had really completed everything I sought to capture in this particular painting style. I got the balance of detail and gestural mark making and atmosphere that represents a personal ideal aesthetic for the medium of oil painting.
But I missed the versatility of a smaller LoRA file, which can be reduced in strength, combined with other LoRAs more effectively, and run with models other than the base SDXL model. So - I ran a LoRA extraction from my own checkpoint model. That thing was nearly as large as an SD1.5 model, but could be resized down to this 200k file, maybe my last ClassiPeintXL LoRA. Nearly exactly capturing the style of the Painter's Checkpoint, but allowing so much greater versatility.
This is by far the most consistent and versatile LoRA for generating images with SDXL that capture the feel of genuine oil paintings for whatever subject matter you want to prompt.
Recommended settings for use:
You can go here (pastebin) to download a ComfyUI workflow like what I used, but without custom nodes that are embedded in my image uploads on CivitAI.
Start with a full 1.0 LoRA strength and adjust down to 0.7 or 0.8 for a subtler painterly effect. You can adjust upward (to 1.2 or maybe a little more) to maximize the painterly appearance, but it can start to introduce some quirks
Use the LoRA with your preferred SDXL model with no refiner. I have so far just stuck with base SDXL1.0 but other finetunes work great as well.
I recommend the DPM samplers, but use your favorite. Some may produce softer painting styles that don't suit my taste as much but whatever you prefer is great.
Don't do anything special for your prompt - just describe what you want to see. I usually start my prompt with 'oil painting of ...' If you skip this, sometimes an image can tend toward more photographic, with blurring backgrounds and a kind of odd mesh between painterly and photographic bokeh. It can be a nice look and you can deliberately get it some by using particular photographer's names. But prompting 'oil painting' keeps everything on target with the LoRAs intended style.
Feel free to add artist names (Klimt, Mucha, etc) and the LoRA will merge its style with theirs with usually fantastic results.
I'm happy to share my work for free, just as the base model SDXL came free to us all. But if you felt 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
The best ClassiPeint yet
FAQ
Comments (7)
Thank you again for all the hard work you put into both painter's checkpoint, and now version 2.0 of this LoRA 🙏.
They are both among my "workhorse" models and LoRAs.
One technical question. Other than the saving in time and electricity, since extracting LoRA from a checkpoint is presumably easier than training one from scratch, is there any other advantages in extracting the LoRA from the check point, rather than just take the same training image set and train the LoRA from scratch?
Always appreciate your supportive words and seeing images you generate - thank you!
I can talk a bit about training, but start with the caveat: I'm the furthest thing from a computer scientist or technical researcher. My background is the fine arts - I'm a painter, who largely has no time to paint, and is currently happy just using AI to satisfy something of that creative impulse.
Primarily, I care about these tools insofar as I can make them work. I read a lot about AI and machine learning, but many things still seem a bit over my head so I just shrug, take what I do understand, and fiddle with things until I get what I want.
I think the biggest limitation on my own training is my hardware. I have 8GB video RAM so I'm only pushing my datasets so far and training so much at one time. I'm also cheap as hell, so I'm not paying for LoRA training. I could try using CivitAI's training, but I started on my local machine and have the ability to easily merge multiple LoRA trainings since they have certain shared parameters. I can't, for instance, pull someone else's LoRA from CivitAI and just merge it with my own without some trickery I haven't bothered to figure out.
So what I've been doing is training datasets of 30-50 images at a time, and then blending them into each other in different strengths. This is very effective and goes pretty far. Then when I decided on a lark to see about merging my various training LoRAs into a full checkpoint model, I found the merging process more forgiving. I can throw more at it, and keep what was essential. So there is actually a little bit of the ParchArt LoRA and PastelArt LoRA and a couple other stray things I never used anywhere else that contributed to that Painter's Checkpoint file. This expands its overall flexibility, but when you prompt for just 'painting' it's staying true to its core style in a way that a merged LoRA of all the same data wouldn't have.
I guess it seems, from my starting point, the size I train my LoRAs at and how I train just one concept at a time, I can't do anything like the same kind of merge that achieves what I got in that Painter's Checkpoint.
I didn't even know that extracting a LoRA from that model would be effective. But my first model extraction seems to coincide with what my eyes tell me - there is way more concept in it than any of my previous LoRAs. When I ran the LoRA model extraction from it, the resultant LoRA was 1.5 GB - far bigger than my 130K file sizes of other LoRAs.
But I also had an instinct that I could run one more operation on that, and resize the 1.5GB LoRA and keep the vast majority of what I valued in it. So I did so and it worked.
Now, is there a route to the same LoRA that I'm calling ClassipeintXL2.0 that doesn't involve this roundabout path, packing into a checkpoint, and extracting back out of it? Maybe? But it's not a simple matter on my hardware. And I'm not presently that motivated to look into it, as I've got what I want
I hope that was enough info for you 😁
@eldritchadam It has always been a pleasure to use your models 🙏, so I am more than happy to contribute to your image galleries 😁
Sorry that it took me a while to reply. Your comment is packed with information, so I had to re-read it a few times to get a better understanding of your training method.
What you've described is fascinating for me to read. I am ashamed to say, I've not created any model myself due to laziness and the simple fact that just keeping up playing with all the wonderful checkpoints and LoRAs people are posting is almost enough of a full time job for me. Good thing I am a retired ex-programmer who can afford to play all day long.
Generative A.I. is such a new and poorly understood field that we are very much at a period similar to the early days of aviation, where even an amateur enthusiast can make contributions to the field. Due to your hardware limitation, you probably have come up with your own unique and novel way of training both LoRAs and checkpoints. I've always assumed that at least 12 GiB is needed for training SDXL. The fact that you can produce high quality LoRAs and the painter's checkpoint using only 8GiB is a testament to your skill as a model creator. I have been very pleased by the flexibility and prompt following characteristics of Painter's checkpoint.
BTW, I am a cheapskate too, and I take perverse pleasure in trying to get away with producing results with minimum resources. While in engineering school, I always use my $20 TI basic calculator, amazing my classmates by keeping up with their fancies $200 HP scientific calculators. I also like to buy trailing edge technology for the best value for the bang. Sometimes constraints force one to come up with good solutions and also making maximum use of one's resources (and yes, I re-use and recycle everything, not just to save money, but also because I just hate any sort of waste).
If you have the time, maybe you can take this comment as a basis for an article for either Reddit or Civitai, as I am sure many modelers and would be modeler will find it very interesting and informative.
Once again, thank you for taking your time to write such an informative reply, it is much appreciated. Your openness and generosity in sharing knowledge and helping others on Reddit had nudged me to share whatever little I know about SD openly without any reservation as well. 🙏
For people looking for a good SDXL model to pair with ClassipeintXL, I would recommend the following based on my own experimentation (ordered alphabetically):
https://civitai.com/models/218300/paradox-2-sd-xl-10
https://civitai.com/models/101055/sd-xl
https://civitai.com/models/120765/sdxlnijispecial-edition
That base SDXL works well is not surprising, since it was build for flexibility, as a "jack of all trades".
Paradox 2 by https://civitai.com/user/Thaevilone is a very flexible model, good for "wild idea" and "crazy artwork", and pairs well with many LoRAs Just remember that photo realism is not its strength. Take a while to get a hang of it, but once you do, you will be going back to it a lot. Please look at the prompt for many of the interesting images in the gallery, and learn from them. Pay special attention to Thaevilone's own excellent, often wild/funny images.
SDXL Niji SE by https://civitai.com/user/Stan_Katayama is an often "underappreciated" model because people immediate associate it with Anime. It is no doubt in my mind that it is the finest SDXL anime model out there, with the ability to generate images with a genuine "anime/manga" feel, better than any model out there. But it also offers flexible, often very dynamic composition due to the anime artwork that undoubtedly went into its DNA. Give it a try by pairing it with ClassipeintXL if you don't like the "anime look" of the model, you just might be pleasantly surprised by the result.
Everything I said also applies to many other excellent style LoRAs out there, special those by EldritchAdam.
If it doesn't work on a specific checkpoint, just change the weight to 1.6. Worked for me.
thanks for the thought. It's not so much that the LoRA doesn't outright work with a few checkpoints - it's that it goes sideways. Colors shift toward harsh orange tones and textures become excessivelt tortured looking. Much less like an oil painting and more like just a psychedelic ugly photoshop filter.
@eldritchadam Thanks for the info. Great work!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.