Bondiment

Hello, it has been a long time since I have played around with Stable diffusion I got into this quite heavily back in 2019 and I intend to continue persueing the entertainment of prompting and image generation. Since I have come back about 2 weeks ago I have been taking a look around at all of the wonderful work many of our peers here have done on CivitAI and thought I would try my hand at some model merging in Comfy as I have only ever merged small Checkpoint or LoRa models back when SD1.5 was booming.

This is a simple model merge between the afformentiond (MiaoMiaoHarem)[/model/934764/miaomiao-harem] and (One obsession)[/model/1318945?modelVersionId=1897112]

It is by and far one of the simplest merges I have done. as it is just

75% of the model base for MiaoMiaoHarem

25% of the model base for oneObsession

and

100% of the clip base for oneObsession

processed via ComfyUI

The reason for the merge is that I found that between these two models there was an absolute degree of freedom throughout prompting so long as the user was very specific. I had found that this model allowed for much longer clip strings than most other merges I have been attempting lately due to OneObsessions hyper specific prompting capabilities and long tokenization process while allowing for much of the data set in MiaoMiaoHarem to be picked up by these extended strings which allowed for avoiding the stalemate that it feels one would get from images generated after writing out a long prompt in OneObsession.

Both of these models are wonderful on their own and frankly better on their own in a lot of ways but I encourage you to hop on in try this one out and let me know what you think!

I will not ask for any support I ask that you show support to the people that built up such extensive models and allowed for a community to play with them. you can find those models to support below

This model works decent in a wide array of samplers and schedulers due to the overlap and range of the two it was merged from. Below Is where I will update what I have found works best for me.

CFG- typically keeping CFG between 2.2-6.5 will generate an image that is recognizable however I have found the best results to be between 3.0 and 5

Sampler and step attestation can get really wonky depending on the length of your prompting capabilities I would say the more specific you get within your prompting guidelines the higher you want to move your steps and cfg regardless of what it is that you are working on.

for best results I utilize the following in order

35 steps on dpm++_2m Karras

40 steps on euler_a exponential

50 steps on heun++2 Karras

(MiaoMiaoHarem)[/model/934764/miaomiao-harem]

and

(One obsession)[/model/1318945?modelVersionId=1897112]

That being said there are a lot of new things to look at and I was definitely overwhelmed and thought I would try my hand at merging two illustrious models together as I had never even heard of these before I got out of playing around on here back on automatic1111. I want to make sure that due dilligence is done and I can be easily confused with all of the new permissions cards that have been developed here in my time away so I went ahead and checked the model host hugging face pages

(Here)[https://huggingface.co/OnomaAIResearch/Illustrious-xl-early-release-v0]

and

[Here](https://freedevproject.org/faipl-1.0-sd/]

As I understand it according to the links above doing a simple merge between these two models should fall under the "Fair AI Public License 1.0-SD". In addition to this I have reached out to the host of (MiaoMiao Harem)[/model/934764/miaomiao-harem] (MIAOKA)[https://civitai.com/user/MIAOKA) who recently published (This tool guid)[https://civitai.com/articles/16163/miaokaimageeffects-a1111-webui] for automatic1111 users. To ask their express permission to maintain the merge.

Thank you for your time.