As I don't want to mess around with lots of different Promt Styles or Loras to get something I want to see, I started a Project aiming to get a dreamy-girls look in it's base state. After a lot of attempts, this is the first result which I liked.

The VAE is baked in. Sucessfully tested with Forge/automatic1111 and comfyui.

Settings I use and recommend:

• CFG: 5-9

• Steps: 24-36 but feel free to experiment with lower steps or higher

• Sampler: euler a or dpm++ 2M

• Scheduler: SGM Uniform

• HighresFix: 1.5x works great (do not use 2x - go for ultimateSD Upscale instead)

My experiment started with the Pony Diffusion V6 Checkpoint. Then I used comfyui for my merging attempts. I did a lot of trail and error, even merged some of my old sdxl Lora attempts and merged some unobtainable models into it. The result can be used like any other Pony Model. If you want a style which is more "flat Anime like" you do not have to use the "score tags".

For me these are working well:

Positive Promt:

score_9, score_8_up, score_7_up,

Negative Promt:

score_4, score_5, score_6, blurry, low res, lowres, low quality, worst quality,

Also use "3d" in negatives if you get results which are too 3D or if you don't like 3D at all.

Another big Influence has "muscular female". I recommend to use the Promting Techniques within Forge/Automatic1111 to use i.e. "muscular female" in negatives after 20% of the image generation:

"[x:number or percentage in 0.x]": Ignores x until number steps have finished. i.e. [muscular female:0.2] will add "muscular female" after 20% of image generation or [muscular female:10] will add "muscular female" after 10 Steps.

"[x::number or percentage in 0.x]": Ignores x after number steps have finished.

"[x:x:number or percentage in 0.x]": Uses the first x until number steps have finished, then uses the second x.

This is how you will not get a overly muscular female but you will get the "scenes" where those muscular females are trained on - which will lead to a bit more dirty stuff. You can do the same with comfy - but a bit more complicated. Therfore you can use one of my images - the workflow is added.

All embedded Artists seem to work well. For the most of them you will have to use a "special code" for these there is a List out there with reverse engineered hashes. Some of them you can find in my example Images.