My first attempt at making/merging a checkpoint all feedback is welcome and appropriated
Please check Version Detail for the changes under each version besides V1 you can find that one below, I don't want to make a mess of the description the way some people do :)
I had an idea what if I used images I created using JAM to make a checkpoint merge, so I did just that I took images I made and trained them on NoobAI Epsilon-pred 1.1 and this is the results, I'm very happy with it please let me know your thoughts as well.
Merge info:
I took the lora and slapped it on top of the NoobAI Epsilon-pred 1.1 (yeah that's it)
My settings:
Sampling method: Euler a, Schedule: SGM Uniform (there's a comparison between 3 different sampling methods and schedule types in the review images)
Sampling steps I go 24, Hires steps 12, CFG: 6, Clip skip: 2
Imagine dimensions for the review images: 832 x 1216
4xNomos8kDAT Upscaler, Upscale by 1.5, Denoising: 0.4
AAA Anime VAE SDXL v2 as VAE (merged in the checkpoint)
uddetailer for faces, I use mediapipe_face_mesh set Dilation to 50 and Y offset to 30 to get the full face
Prompt:
masterpiece, best quality (those should be the bare minimum)
masterpiece, best quality, newest, realistic, photorealistic, hdr (What I use/recommend, there's an image comparison to see if realistic, photorealistic has any effect or not on the image)
Negative prompt:
lowres, worst quality, low quality (those should be the bare minimum)
lowres, worst quality, low quality, bad anatomy, worst aesthetic, jpeg artifacts, scan artifacts, compression artifacts, old, early, loli, child, bokeh, distorted anatomy, bad proportions, missing body part, missing limb, unclear eyes, bad hands, mutated hands, fused fingers, fewer digits, extra digits, extra arms, missing arm, missing leg (What I use)
Checkpoint info:
NoobAI function:
As this is merged on top on NoobAI it should retain concepts and characters that you would normally get when using just the NoobAI Epsilon-pred 1.1, and the same goes for all the bad.
Generating women vs men:
Because all of the training images I used were all women this model tends to generate better women than men.
Thick/muscler:
Again because of the training images used the model will tend to generate thick/muscler women even you don't prompt it this is by design yes, there's an image comparison to see the difference for when you prompt and when you don't.
Random black lips:
same thing as thick/muscle because of the training images, black lips tend to be generated even when you don't prompt them so look out for that.
Random futanari:
I don't why it sometimes it, now I know what you're thinking it's got to be the training data but there where only 5 images so it could be the model is just horny.
credits:
@Purfecter for creating JAM in the first place and letting use images produced my it
@L_A_X for creating NoobAI Epsilon-pred 1.1
@bluvoll for creating Initium V2
@Manityro for creating toml used for the merge on ComfyUI
@Everyone at the Arc en Ciel not only they made every lora before and after this one better but for helping my dumbass with all the questions as well
I'm sure I'm missing someone so thank you too
feel free to make your merge using the checkpoint