The checkpoint type is set to "Trained" because Version 3.0 has been finetuned. Version 2.0 and 1.0 were purely done by merging checkpoints though.

The baked in VAE is not in fp16 precision, so trying to load it with --vae-in-16fp will produce black images. If you want to use that feature, I would recommend the modified sdxl vae from https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main. The decoded images will be different, but the difference should only be minor.

Version 3.0

A finetune on top of a Merge of NoobAI_epsilonPred10 with Noobieaterv20. Softens out the harsh linework of the previous versions, while keeping the more saturated colors and higher amount of detail.

V3 was done in two stages. First, a similar approach as V2, where I merged NoobAI_v1.0 with Noobieater_v2.0 through the perpendicular component method, while keeping the NoobAI CLIP. I tried to enhance this merge by adding small amounts of other models, but found these to end up worse, so I stuck with the initial merge. That model was then finetuned for 3 epochs on a small dataset of curated artists and concepts (~1000 images). My goal wasn't to change a lot, but just to soften the Itercomp influence on the base model and I feel like it achieved that, and it is actually more aligned with BaseNoob in terms of composition and varience than the base Itercomp merge was. It's strange, but I'm happy it worked out that way.

Version 2.0

A merge based on NoobAI-XL (epsilon-pred 0.75-version). While losing some of the creativity in posing, it is still rather close and gains stronger colors as well as improved character coherency and even gets more consistent at generating hands and other details.V2 was produced by trying several similar approaches as V1, such as adding the difference between Itercomp and SDXL to it, as well as the difference between V1 and NoobAIXL_Epsilon-pred 0.50-Version, and then merging the best versions with each other and the 0.75 version itself. I repeated this step a few times until I felt like any additional changes only worsened the outcome. I kept the 0.75 Text Encoder, with the goal of preserving the increase in characters it knows, and I feel like not only did it keep its knowledge, but stabilize it a bit as well.

The goal was again to retain the character knowledge, pose variety and general feeling of the base model, while enhancing it as much as possible, especially the colors. While I don't think I wholly succeeded in this, particularly in keeping the poses as varied as they were in 0.75, it's as close as I can get it while keeping the colors the way they are now.

Version 1.0

A stupid little merge based on NoobAI-XL (epsilon-pred 0.5-version). It is a bit worse with some artist styles, however I like the colors and feel like the coherency has been improved, so I am sharing it. I would classify it as a general sidegrade that's useful if you want more saturated colors.

To get here, I merged Noob with the perpendicular component of IterComp (https://huggingface.co/comin/IterComp), and then merged that with a blockwise merge of Artiwaifu Diffusion 2.0 and Illustrious. I kept the original NoobAI Clip and used a slightly altered VAE that I found on an anonymous mongolian basketweaving forum.

Thank you to reoe for the perpendicular component idea