The LAION-5b dataset, which SD 1.5 is trained on contains already a significant number of images created in the 1700s by Giovanni Battista Piranesi:
however, their resolution, tagging and chopping is all over the place, this model reinforces the concepts and art style based on high quality 1024 full-sized images without borders text and descriptions from the original artworks of Giovanni Battista Piranesi. they are based on high quality scans by the University of Tokyo library in the Kamei Collection. the images have been properly tagged and perfectly chopped as well as edited to remove numbers and comments and text embedded in the images.
The resulting checkpoint is now available in its first iteration. I’m also uploading the training data if you like to pick up where I stopped. LoRa
Description
The LAION-5b dataset, which SD 1.5 is trained on contains already a significant number of images created in the 1700s by Giovanni Battista Piranesi:
however, their resolution, tagging and chopping is all over the place, this model reinforces the concepts and art style based on high quality 1024 full-sized images without borders text and descriptions from the original artworks of Giovanni Battista Piranesi. they are based on high quality scans by the University of Tokyo library in the Kamei Collection. the images have been properly tagged and perfectly chopped as well as edited to remove numbers and comments and text embedded in the images.
The resulting checkpoint is now available in its first iteration. I’m also uploading the training data if you like to pick up where I stopped. LoRa version is also coming.
FAQ
Comments (7)
This would be very useful as a lora
is a LoRa to see https://civitai.com/models/16428/giovanni-battista-piranesi-1700s
Comparing this with the Lora, I know i'd be able to use the lora with a variety of other models, but is there a specific advantage to using the model over the lora? Or the Lora over the model?
Thank you, this immediately caught my eye and really want to sink my teeth into it, just need some data to help myself decide.
The LoRa extraction process and the process of Reintegrating LoRa introduces floating-point imperfections concerning the bit depth of your checkpoint 8 16 FP 16 BF 32 64 and so on. so if you want to have the full quality, you could go with the checkpoint if you like to mix and match it anyway don't care that much about the exact representation and are okay with an approximation then the LoRa will work just nicely. we're talking here about a maybe 1 to 5% degeneration in data to final image quality.
@ClassicRPGArt thank you for the in-depth explanation, definitely going with the model :)
@ClassicRPGArt You are doing God's Work ! Thank you so much !
An angel has blessed us with a Piranesi model. As though my mind were read.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.