TLDR: USE THE FOLLOWING MODELS IF YOU WANT GOOD RESULTS FOR HOLOS: bbw-hll3.1 Final, bbw-hll3.1b2, bbw-hll4b2, bbw120kg 6.0+1.0 You Can (not) Lose Weight, USE THE FOLLOWING IF YOU WANT ANIME OR YOU WANT TO MAKE MERGES: bbw_step_dep3-chilloutMix-half, bbw_dep3co_Cocoa-half, bbw_dp3coc_Tea-half
These models are a merge involving bbw-30000k-step.cpkt in order to resolve the issues with them. I am not the original trainer of the bbw-step-30000k.cpkt model. However, I am the creator of the Cocoa and Tea merges. The Tea model builds on the Cocoa model by merging Tea into it. I will include the original post I made regarding the models below. Additionally, I have included two beta hll merges, I personally prefer bbw-hll3.1b1 over bbw-AVL3.1, the base model for both of these is bbw-tea-half.
(If you want to prompt {{{Bigger}}} holos use bbw-hll3.1b2 or bbw-hll4-b2 as their b1 counterparts both have a ruined clip encoder which was force fixed in b2.)
https://mega.nz/folder/v80QhCCZ#TaTb_8vErCq0efZYdFh35Q
(Most of the models here in the above link need to be pruned.)
Some general notes regarding the model: The model is based on Anything V4.5, which means that this should be compatible with most LoRAs and embeddings. If you want to get backgrounds that are even more detailed, using the token "detailed background" will activate dpep3 much more aggressively. As for negative prompts, go for whatever you would use for an Anything v4.5 or AOM2 based model. Some notes regarding the Cocoa and Tea Merges: These two have some oddities introduced from Cocoa which causes the masterpiece keyword to create some weird results, although not consistently, so experiment with it in your prompt and with it not in your prompt. I have had good results with and without it. Additionally, make sure your clip skip is set to 2, otherwise you might encounter some oddities.
Recipe for the merged model First, merge dpep3-chillout into bbw_step_30000k (when I say this, you should make bbw_step model A, and dpep model B) (find it here: https://huggingface.co/closertodeath/ctdmixes/blob/main/dpep3-chillout.safetensors ) with the following Merge Block Weights (you cannot use the default checkpoint merger, use supermerger, or another extension that can do MBW) 1,0.9,0.7,0.5,0.3,0.1,0.4,0.4,1,1,1,1,0,0,0,0,0,0,0,0.1,0.3,0.5,0.7,0.9,1 (You can start from the half version in the above link if you want, the fp16 and normal version did not have any hash differences) Then merge the previous model you just made, I would call it bbw-dpep3-chillout, with Cocoa at the following block weights: 0.15,0.15,0.15,0.15,0.15,0.31,0.3,0.3,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15,0.15 (NOTE: FOR ANY MERGE INVOLVING SPECIFICALLY THIS MERGED MODEL, ONLY DO MERGES WITH THE FULL SIZED MODEL AS THE HALF SIZED ONE HAS A DIFFERENT HASH FROM THE FULL SIZED ONE) Then merge the previous model you just made, I would call it bbw-dp3c-Cocoa, with Tea at the following block weights: 0.35,0,0,0,0,0,0.1,0.1,0.1,0,0,0,0,0,0,0,0,0,0,0,0,0.35,0.35,0.35,0.35
(You can find Cocoa and Tea here: https://huggingface.co/andite/desserts/tree/main Cocoa is most likely a mix of YohanMix and Blossom-extract, but I have very little information on Tea )
Anyways, let me know what you think, I feel like the merge managed to really fix a lot of the issues the base model had. My favorite is probably the tea merge. HLL Models are in the repo.
Special Thanks:
/vtai/ - for getting me back into proompting
/bbwai/ - for piqing my interest with bbw-30000k-step model
anon - for making and training the bbw-30000k-step model
hll anon - for making and maintaining the hll model and LoRA
andite - for making the desert models which really did help fix the model
closertodeath - for making ctd and it's model mixes which fixed the backgrounds
Aotsuyu - for the base recipe that I used to get to the Tea Merge.