Remarkably faithful to the original SDXL model, ParchArt XL. The primary motivation for this model is the substrate: parchment. Most AI, when asked for parchment, generate something that looks usually like a toothless paper stained with tea. I really want something that makes me want to touch and feel it. The annotations and illustration style are secondary elements to this.
V1.2 is trained on a larger base dataset than 1.0 - which has always been a purely synthetic dataset, starting with some old Midjourny v3 images. But this one merges (to a small percentage) a model trained on IRL parchment illustrations and illuminated manuscripts, which helps Eldritch Parchment to achieve a bit more color saturation (if asked for) and coherency in drawing, and legible rendering of explicitly-prompted-for text (i.e., not a block of text, but a bold title).
I have toyed with this model more than any other (even if my oil painting model is a little more dear to me) but this is definitely the best it's been. It follows prompts extremely well. It will give you loads of annotation and scrawled text when you ask for it, but mostly only if you ask for it - prompt 'annotated' or otherwise describe the text blocks you want if you are looking for that. And it can shift between being more illustration-focused or parchment-texture focused really easily based on your prompt.
I won't say I couldn't be happier. I can still imagine it being a little better. But it's pretty darn good!
Description
initial version - with likely successors, but probably not too soon.
FAQ
Comments (8)
wow bro very cool,indeed flux has opened new doors :)
Thanks! So many datasets to use for Flux training. It responds remarkably similarly to how SDXL did to the same datasets. Which is fantastic, because you can anticipate results, but have much better anatomy and prompt comprehension. Though this particular model really flattens the potential for complex scenes. If you aim to get that parchment look at all, the prompt needs to stay fairly simple.
Very nice Adam, useful at lots of different strengths.
Thanks Rob! I always love to see what happens with any LoRA in your masterful hands.
I don't know how you made this lora but it is almost spookily useful. At low levels it improves almost any generation. I wondered what the same approach to other surfaces might bring, plaster for fresco... or even surfaces you might not paint on in reality stone and wood etc.
Hey Rob - it's a weird one to be sure. Everything about this model is synthetic data. Started with some old Midjourney v3 images prompting for maps and parchment things ... that old MJ has some great textures and I wanted to capture that. So I first trained an embedding for SD2.1 and that thing was a weird chaos machine. Got a lot of that parchment and ink scrawl look from the original material, but was just wildly unpredictable too.
When I started training for SDXL, I saw quickly that the same training data was being understood quite differently, and allowed me to prompt for nearly coherent illustrations. So I took its initial output, ran generations with partial strength of my oil painting LoRA to add coherency (it lends to a decent organic texture too) and created from that combo a new set of training data ... did this sort of thing a few times to try and find a good dataset. Manually edited some of the images and redrew a thing or two here and there, retrained etc.
The last dataset that I used for SDXL I used for this one too for Flux.
Fascinating, so Flux sees the same training data differently to SDXL. I didn't realise there was so much try and try again. I'm finding with flux loras that many cause coherency issues where bad anatomy and five fingered hands return, parchment does almost the opposite. Another lora which does this is StencilArt, flux is a strange beast!
@eldritchadam I always enjoy reading about how people made their models. Thank you again for sharing your insights 🙏
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.